diff --git a/.gitattributes b/.gitattributes
index bb30d900d34f64e96ca5fce309a246c5457b938b..6a6bb86827b428d4b1fb8dc49fa6844d9a77e3ee 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -122,3 +122,7 @@ inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/equivalence_e
 inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/equivalence_easy__run_20260409_134704__pair_000778_B_traj_000779.json filter=lfs diff=lfs merge=lfs -text
 inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/equivalence_easy__run_20260414_132732__pair_000596_A_traj_000596.json filter=lfs diff=lfs merge=lfs -text
 inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/equivalence_easy__run_20260414_132732__pair_000596_B_traj_000597.json filter=lfs diff=lfs merge=lfs -text
+inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_easy__run_20260422_172644__traj_002919.json filter=lfs diff=lfs merge=lfs -text
+inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_hard__run_20260427_001012__traj_000047.json filter=lfs diff=lfs merge=lfs -text
+inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_easy__run_20260409_134704__traj_000435.json filter=lfs diff=lfs merge=lfs -text
+inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_hard__run_20260427_001012__traj_000451.json filter=lfs diff=lfs merge=lfs -text
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_easy__run_20260422_172644__traj_002919.json b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_easy__run_20260422_172644__traj_002919.json
new file mode 100644
index 0000000000000000000000000000000000000000..b60a6b20ec4bc46e3e373aa228b57ceeebcb95c0
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_easy__run_20260422_172644__traj_002919.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1570975ea5671ed4b8fb2f6bfc5223724dee7023608e0954bc95cfb80b38596
+size 33587868
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_hard__run_20260427_001012__traj_000047.json b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_hard__run_20260427_001012__traj_000047.json
new file mode 100644
index 0000000000000000000000000000000000000000..e206847e671db78aa36447846aaa6ca21dd6a835
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/inverse_hard__run_20260427_001012__traj_000047.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac0fed8a3b301bb2b7c7d282219593edeaaadf3b6ec84a377ec8a8f3051999f7
+size 33560770
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_easy__run_20260409_134704__traj_000435.json b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_easy__run_20260409_134704__traj_000435.json
new file mode 100644
index 0000000000000000000000000000000000000000..84439ba07c242069fb1f39ee63266bfece1b26bb
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_easy__run_20260409_134704__traj_000435.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81ba89c7f5989a072b16029fcb135b814f3cdb66248e79d02006386fc34b02f1
+size 33549237
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_hard__run_20260427_001012__traj_000451.json b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_hard__run_20260427_001012__traj_000451.json
new file mode 100644
index 0000000000000000000000000000000000000000..ef87745397032fdbba9b80963b9dbc8517a5dc97
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/attention/loop_hard__run_20260427_001012__traj_000451.json
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:105cbac19ce3fe78071f1351e66288ee742675f1f035d2b11e2acace28043d0a
+size 33588647
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144460642845/inverse_hard/run_20260427_001012__traj_000047/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144460642845/inverse_hard/run_20260427_001012__traj_000047/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1b1b438316f230aa6d9f6012a816b2df4bcef029
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144460642845/inverse_hard/run_20260427_001012__traj_000047/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eac1f324170e2c5ac92d235beadf2eceaf42993876554a2f51fc8f8ef40fd945
+size 5512714
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144619641557/loop_hard/run_20260427_001012__traj_000451/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144619641557/loop_hard/run_20260427_001012__traj_000451/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..7c6304089e4286bd809e6182ab2b0e4269995a06
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144619641557/loop_hard/run_20260427_001012__traj_000451/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc67f65eb099e12220df7b6a0b91cfad8d623faa959807200f292cb75fd3af13
+size 10723208
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144760682356/equivalence_easy/run_20260409_134704__pair_000778_A_traj_000778/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144760682356/equivalence_easy/run_20260409_134704__pair_000778_A_traj_000778/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e2458c50269cc06f6546e221be6e0ea4f4714718
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144760682356/equivalence_easy/run_20260409_134704__pair_000778_A_traj_000778/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fcbd3b53f6816c5bb77abf7188738b7176a865a4c1b6c8fa33f88a8e3b504cc
+size 29786989
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144861177561/equivalence_easy/run_20260409_134704__pair_000778_B_traj_000779/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144861177561/equivalence_easy/run_20260409_134704__pair_000778_B_traj_000779/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cee8bfea842e423b99764a8964ac653ea0e99268
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/failed_attempts/1790103144861177561/equivalence_easy/run_20260409_134704__pair_000778_B_traj_000779/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79cc35e3e8b1f9cacc2a3af97e15be59e6f50d3e1021e53ac7d0506bdbdde526
+size 29301408
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_000047_A_traj_000047/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_000047_A_traj_000047/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6387349640a7483a9c46f7e07c10794f5b3f3635
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_000047_A_traj_000047/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46b1fbc50edbd06f4f6393714cf21107fda8b74bd76138eec06ed3aabe8e1870
+size 25146056
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_000047_B_traj_000048/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_000047_B_traj_000048/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..61a3d317c5483bae1949f4421d110ee4088722e2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_000047_B_traj_000048/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:124f4629bb5f00504c7b39a24860bd5f7349e5339242153d63f05e1388153545
+size 20898927
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_001790_A_traj_001790/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_001790_A_traj_001790/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4775db80bcafa6c64899bdb31d3d573b4666bf41
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_001790_A_traj_001790/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:044b3f3424622934817b310ce52a8266ac13eb36148cd43e1ca02444fd72e8e1
+size 15510493
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_001790_B_traj_001791/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_001790_B_traj_001791/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..49c5ab13cb3c1e5046c058a76f5b4fec500b9fdd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260407_174321__pair_001790_B_traj_001791/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77dd773a265468e1f4a779b6867595ca6848cb37fb13710d7fe7a4aa66c07c32
+size 15839627
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000476_A_traj_000476/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000476_A_traj_000476/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..404167e1b268a319fda11a32fb3095b7223b41dd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000476_A_traj_000476/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5921555762ad44dbf363307d7516e7d9e55eda90388c2f199ef89726fc04616
+size 34672950
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000476_B_traj_000477/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000476_B_traj_000477/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..86a623c07a92dac46437a45101518f90ae9c7212
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000476_B_traj_000477/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfc6c017bc86f9cd31f449f15e54c2a40fae13c7f01fcd22589804e014b9c2be
+size 34953965
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000920_A_traj_000920/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000920_A_traj_000920/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e7ae02e26b7a90167d3864dd5c0de46233dbe8bb
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000920_A_traj_000920/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5c350187031f098e9ef997d6dc79269f9d6586ba4a50b39a510b4b497eadc3b
+size 23649716
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000920_B_traj_000921/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000920_B_traj_000921/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..414a1096104c61f620e55f9742addbb5ab797949
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_000744__pair_000920_B_traj_000921/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:092a969fb6f4ac54dfc84e4eee29e976d03e677d2f2d7bf9806ac8c21ba05d99
+size 24343860
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002128_A_traj_002128/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002128_A_traj_002128/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cc8b5f72611230f526e3e19963f3902e43d01683
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002128_A_traj_002128/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2a8e338ff7683b625605516ec744f4acf8cd43902d5670dae5f2e5a84e951ac
+size 17982802
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002128_B_traj_002129/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002128_B_traj_002129/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..a057ea4bcad36e18d3b4eb6d5b1a30c084780b8d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002128_B_traj_002129/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bb61ec65132ab130737e6db9ff045e76764180586edea3bc89fec0f995e362c
+size 28815412
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002320_A_traj_002320/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002320_A_traj_002320/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..697f52d3631cb4427175b1006378eb4bff539ecb
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002320_A_traj_002320/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64a76123bd3fa3bbd4e7e0c7686368622507609b9278793d82db855487fbe9f6
+size 30573789
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002320_B_traj_002321/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002320_B_traj_002321/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6bf5c600568bdbd163a3ad781ccf70af12716e18
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_002320_B_traj_002321/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e82d73c2438ff67ae402e660adc4c93ca8a14e0d7c4f3beba6aece335d449f65
+size 27565962
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004149_A_traj_004149/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004149_A_traj_004149/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..2746dac7c03496c255fe0fe72e6c2d72ddf0a91d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004149_A_traj_004149/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a8675b23672a5fa2f0075d9b8e4d68394c6ddedd837b5359a567f46a8cf75d8
+size 23546232
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004149_B_traj_004150/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004149_B_traj_004150/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..86958e11b9f9ea273548466b2d7c8f70ada7e22b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004149_B_traj_004150/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9734917c24cd9a68d407828d8936faa28ded709eb52d3e56608fc0d5022245a5
+size 26421539
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004865_A_traj_004865/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004865_A_traj_004865/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b4673a9d1566a72ff7761cb0877d73f136774972
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004865_A_traj_004865/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7f4c1938b08ae07a7e9120159c7f6428e97783123ffa3213ded5b05869e4f06
+size 22580740
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004865_B_traj_004866/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004865_B_traj_004866/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ce0e89bc0ca24d1ff4926921889fa20685002331
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_004865_B_traj_004866/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26524c2cfe90b1eaad584c81f9c1b8dab94154bc886bc8a8e072fc5648369dad
+size 26998136
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_005916_A_traj_005916/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_005916_A_traj_005916/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..aa2441beda3540a20b2b474e9b5b3c06b95806dc
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_005916_A_traj_005916/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92380ea1fc2d8609a3417cadfd50da68e71686e56691590e242623349c9617b5
+size 19347753
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_005916_B_traj_005917/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_005916_B_traj_005917/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..88c660e4eb4a842b2a54e43af4286159f53ebf37
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_005916_B_traj_005917/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8baf8531127774083b56d2ef380d763c5bd76d060cb8d282a9f8314dea913b6
+size 18593804
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_006244_A_traj_006244/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_006244_A_traj_006244/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..22bddadb4c3617d63b811849015ba2e47b007ac4
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_006244_A_traj_006244/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96681d48e0aead9d6394791b9cae623c7c26a25d461d5193daf4dfab092e06ea
+size 24814470
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_006244_B_traj_006245/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_006244_B_traj_006245/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..83f7dfc4d1d6dd0336b78fb2589aad656d4dabe2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260408_180248__pair_006244_B_traj_006245/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a58ac04ee3935e15eadac438b7fac9e75f626699253a1dbcdfb41b0bd13f531
+size 30242982
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000315_A_traj_000315/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000315_A_traj_000315/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d7545f2fe4c18149a5117522ee7e2b444f2a6491
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000315_A_traj_000315/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98749f1646d708de562c496e406c3847fc17fb6438fb25c85097e60e58b0fd41
+size 16283561
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000315_B_traj_000316/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000315_B_traj_000316/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e92c62298ba96116bef4c918338b7ac5912aff1c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000315_B_traj_000316/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf69610665319025032e2780638fba899d56058f1b27d34c9c99ae08df458394
+size 14202757
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000657_A_traj_000657/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000657_A_traj_000657/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..2560b8419093001e9d88e5e33f309d25e67f9972
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000657_A_traj_000657/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:983f1ca49ea29db46ec3ae7e6f314c428ccab92e1c391e199bac3ec71f4fc911
+size 37213362
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000657_B_traj_000658/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000657_B_traj_000658/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..3d1ce464816b1500ccdf6de51231aa4b40fbec2f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000657_B_traj_000658/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55f35dcaa60a7bf17f8cd24dfaf2a5b090b4e18f3c0100ad2b362697dc200c4b
+size 30826412
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000778_A_traj_000778/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000778_A_traj_000778/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e2458c50269cc06f6546e221be6e0ea4f4714718
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000778_A_traj_000778/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fcbd3b53f6816c5bb77abf7188738b7176a865a4c1b6c8fa33f88a8e3b504cc
+size 29786989
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000778_B_traj_000779/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000778_B_traj_000779/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cee8bfea842e423b99764a8964ac653ea0e99268
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260409_134704__pair_000778_B_traj_000779/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79cc35e3e8b1f9cacc2a3af97e15be59e6f50d3e1021e53ac7d0506bdbdde526
+size 29301408
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000166_A_traj_000166/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000166_A_traj_000166/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..72e2b0e55c300cc5679a9374ca3ccdb801e53d51
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000166_A_traj_000166/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:345d528b3b0870784b4877066c15aba59f680c5b52916c125776d06d8b5fb39d
+size 20157773
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000166_B_traj_000167/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000166_B_traj_000167/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..00338ebba687639a579614968e1a4ae752d19512
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000166_B_traj_000167/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c4ccc1a1dd305a7fccf03c4751cd24775287f73e5364b6ec4f03e775ddd552b
+size 19992000
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000210_A_traj_000210/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000210_A_traj_000210/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..f7c544152b3cd02ded44747a24663542410a7560
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000210_A_traj_000210/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23e95939b7bcacf25778434f9174fe04e06bd8b193300fd9cecd7fa2a80fced5
+size 26689165
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000210_B_traj_000211/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000210_B_traj_000211/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..f6576f1973ab9795acdb3addb765977dcc86b871
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000210_B_traj_000211/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8cf789ff25f12b4a021685431274a8f6a5542e64057b4afdcb2d942afa443f3
+size 25627176
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000368_A_traj_000368/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000368_A_traj_000368/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9668f87b8044dd79d14d81e2d524fc016d6a11e5
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000368_A_traj_000368/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88e5f78ab3721d2f6612e2699536b791555754cd458a9c6a8676e3360263af72
+size 19482036
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000368_B_traj_000369/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000368_B_traj_000369/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..68a164b6c35a4cac95fff168e26d86c55ec5daaa
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_112607__pair_000368_B_traj_000369/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9d31816fcc73c53de19c67fec9b42a90dcf7b0a7d45434e4037011056cb09cf
+size 20195494
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_002375_A_traj_002375/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_002375_A_traj_002375/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ffa10436ca176f6a12db8548f1cb2915ee48258b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_002375_A_traj_002375/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f1550b50210d96cedefb2a17539f408e71e13324c4a0a97d2a0152d8e1adcca
+size 34905027
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_002375_B_traj_002376/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_002375_B_traj_002376/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..54e1365fba1d305ad4bd6337d410b4d7ae90f2b8
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_002375_B_traj_002376/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cd242ec5e2aad9188dbecfa09dc341de8a3be0a3d3821c445252862ab8c86b2
+size 32846027
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003641_A_traj_003641/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003641_A_traj_003641/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1f1d2d5ba04859e04737734256021b8566074bc2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003641_A_traj_003641/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4278bf74c103a4087a082f59968e0212ac450785f835016d0a6bd378903a9d2
+size 23322475
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003641_B_traj_003642/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003641_B_traj_003642/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b746e2d8503e9e9296b033fdaace5e619e25ec50
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003641_B_traj_003642/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4928b961597801e2d9433deda028fa1b3382f42c500a5ef53123b23e4f85ebc
+size 23101604
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003864_A_traj_003864/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003864_A_traj_003864/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..87fa6360ff1dec08714a3b45b06598ac9f7c0c00
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003864_A_traj_003864/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f40b973af518501fa02102edfdf1e116d0eaf633b29a1dccf374524c4c78a392
+size 19424848
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003864_B_traj_003865/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003864_B_traj_003865/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..40e2adcd4487131b8ababd418c3aa06942c8a68b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_003864_B_traj_003865/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48ff1ee0a808881a949f8960c5d0f84f6da1eb6614d0c41feaa34f512ba01da4
+size 19426401
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004238_A_traj_004238/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004238_A_traj_004238/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cce5efd98acae22bf5b0ff043776d4576ea74713
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004238_A_traj_004238/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c60e7fd09eb37fc97621b5585358f9ddc44323f10868ca3f3b172242f85a74fc
+size 22150734
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004238_B_traj_004239/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004238_B_traj_004239/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e87d652a9923eb07a33b381a7d917585fa69f40b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004238_B_traj_004239/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf9273ad8b4527534944837afed7583318f7d7a90b87988dee0512628fc9a7f5
+size 23442645
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004544_A_traj_004544/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004544_A_traj_004544/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8eca288a5f5aee7548a7e01ca234cebf4e43e683
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004544_A_traj_004544/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffbf2efd16c96ac87e39348bf9dcf388880ee9e629e4120da3d838aeddcb8934
+size 15486775
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004544_B_traj_004545/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004544_B_traj_004545/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8461bfcb5d0a4fbf534afd91f83b6114d1308a13
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_004544_B_traj_004545/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7e22ca05d7f095244c7c90949348f99d0501940be403e38daa801f6165db58c
+size 12847772
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_005031_A_traj_005031/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_005031_A_traj_005031/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..22f80be8733b86acd9ac4c928f5d52f9ab305166
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_005031_A_traj_005031/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:774d02cdb9bce3792201c160645c60611e49bd2326ff090a00f88db2ad1e2807
+size 32258521
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_005031_B_traj_005032/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_005031_B_traj_005032/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1a3d0d9e84f077cf1ab73f6e843729f8b4792e3d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260410_184831__pair_005031_B_traj_005032/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b548ad1a6ff48d01ac551c82b937e98d20dc36ebd4db5b19102bfdcbe6246f35
+size 32058277
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000020_A_traj_000020/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000020_A_traj_000020/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..5732bccc4ad28186b79df982765301f67f931693
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000020_A_traj_000020/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f43aa23a7a2ff7011aba81d9bec4c3b775fce4f8d3e9ab687c52808180ff0f95
+size 23953961
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000020_B_traj_000021/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000020_B_traj_000021/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..28bc67b1c00c9b74224c49c891230dba075a8b55
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000020_B_traj_000021/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1298906befc8d38f23cf2fc4ce8787a396131b5fbbe240cebb90bd58f9748317
+size 23850843
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000896_A_traj_000896/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000896_A_traj_000896/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d1591e055cc4bb0b20a087bdcb823bef34c5b18b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000896_A_traj_000896/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c58ad90fa95800516c4490715df5d0ea494b17d670b77173a1a953ea4e1c9999
+size 35392280
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000896_B_traj_000897/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000896_B_traj_000897/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1567908583c6f95bcc95c1fd80f09116a87fafe1
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000896_B_traj_000897/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0ec2cf0ed7e58a7a1f7d52d8704a8102dca4a662be5f6c237d1f2665192af63
+size 22275925
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000975_A_traj_000975/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000975_A_traj_000975/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..910dd3a0094ed1b31d879ecf034215b32ec7bb8d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000975_A_traj_000975/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8b9e25c7d7761831dfe987707333e74e7e3cb421bb8840d275ec4cb9d522d0c
+size 15447288
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000975_B_traj_000976/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000975_B_traj_000976/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..36d437cd11080dfee2b3a3a5ac4fc6f86554cb9a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_000975_B_traj_000976/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96414bc9959abb608a7a0342fe733c50a6bec089a93c8a4db7c3b25a42e3cc55
+size 20358636
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001130_A_traj_001130/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001130_A_traj_001130/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..fdf21a1689737388759ea24d0ba141a9c6a24289
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001130_A_traj_001130/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7b37e7105261be7f883d0cb2be8f7b34856aa26ebec5bb13cb7ecef65bcca50
+size 16289595
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001130_B_traj_001131/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001130_B_traj_001131/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..129c46ff4fd28995b9b5ad203a5bef80ad421116
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001130_B_traj_001131/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:348f557a5e56f6bc39334e471019ed4e28ade0334799fe14bd42d1013228cfe6
+size 20199497
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001156_A_traj_001156/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001156_A_traj_001156/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8e40f592890cf77ca5ed349aff70295f029ae139
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001156_A_traj_001156/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a285324b4711fa3d2064a05f1cbe49c589e562f302a742e1d07d8bb93ebc6848
+size 18260347
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001156_B_traj_001157/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001156_B_traj_001157/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b95f0a2d8e8ab8ad1924d70797959d89164634c5
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001156_B_traj_001157/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94bdd73c7273bc8641f997e7841742180cd68883abf53d9a526d9f34c2c8b18b
+size 17609839
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001293_A_traj_001293/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001293_A_traj_001293/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..22c7286be2be5c25f57cf634a2bda0b58b6e4059
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001293_A_traj_001293/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67de48c67efa5ce7c0327046bc4a487527f33a7fbbc7b0dd7ba04ea75b06ad0f
+size 28564090
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001293_B_traj_001294/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001293_B_traj_001294/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..bb1c01e1c6d4dc34a866310634e9cfba606c49f3
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001293_B_traj_001294/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a95b525cc4ea9903e1e1a6ccec06b17f6831d48ce2f8c9c0a11c63c6fd3b8ecb
+size 30757346
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001428_A_traj_001428/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001428_A_traj_001428/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ff3134a076840dc7f2818dffdbecb8b8369b5cf7
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001428_A_traj_001428/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6b436b4e8e8e1af6345cdc6f5b9026d422f91a1c0450c1a552c8f5d80170535
+size 22323796
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001428_B_traj_001429/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001428_B_traj_001429/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..fbf711d97fbf8f2ec8317301766fb36402dd59b4
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_001428_B_traj_001429/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:604ea2264918e8bedbd0fd8dd8154892374e14ad39b526d1fb06acc1183ae354
+size 21783991
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002019_A_traj_002019/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002019_A_traj_002019/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0c36b70c1fd4c1e3803513d9f4ec462ef168b064
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002019_A_traj_002019/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e101a0234e46cf3edfacfe3c6d847bdfb3b0a54a9e8ac992d106fc1fbfe23255
+size 19378195
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002019_B_traj_002020/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002019_B_traj_002020/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..fa0a1e028e21f4ed607e21c7ae04edbd0fa49ce2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002019_B_traj_002020/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddeca632960f8ab97c5429c9202d19132f22d72f040668a59cb17dbe8b723cc2
+size 28493980
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002170_A_traj_002170/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002170_A_traj_002170/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..10f1d721b573a77f720e9a6775af66ec70e06070
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002170_A_traj_002170/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6cc58eaa2acdd1f1b4577360698eba9ab14540d0aec2142fc4cfa50835d057c
+size 18111479
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002170_B_traj_002171/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002170_B_traj_002171/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8a515bfe988b2f0aeaa3c6dbc17a395ff9f72021
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260411_181558__pair_002170_B_traj_002171/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc3c7cf19ae30d6f4b4641bd9e3d38d72274f8b19991af69f3130455ae529fc6
+size 16348978
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000294_A_traj_000294/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000294_A_traj_000294/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..83dd7a17ed5e4ebc24b23d8a1d145c7e6f0bff77
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000294_A_traj_000294/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1708033ea1863aa3981c3d110fdfc16ecf1055f7cced951978791afa1b095a69
+size 9938593
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000294_B_traj_000295/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000294_B_traj_000295/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0011b8a711283eb6f441b6c119b4ad821d05d585
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000294_B_traj_000295/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3af7aa8dcc91cb2561f0f4389e7b190ee25bb80cd58d44680954b7006cfaec2
+size 10342218
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000698_A_traj_000698/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000698_A_traj_000698/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..82440977bb6e6e8a34cc97a1f1977c00b2f2bca0
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000698_A_traj_000698/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbd591363ab5874282101d6864cbfdd821373329d29a8552ec5620ac8ec2431e
+size 21112767
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000698_B_traj_000699/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000698_B_traj_000699/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e88e6632f70960d09528f918867c8a27eadcc5b1
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_000698_B_traj_000699/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdda1e5f4cbf0e19e4742140aa0ba5a2b5c8f433a38ed19663c8db1ad4d9160e
+size 25327466
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_001596_A_traj_001596/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_001596_A_traj_001596/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4af514c076594716d5943776996ecef648aa9520
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_001596_A_traj_001596/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8c7a91709917bfc9cac25fd2591d681307a5bd45ab58026047b68cb1f201160
+size 9321128
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_001596_B_traj_001597/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_001596_B_traj_001597/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..786fc866a22a197de9279aa150eda7a3f57725d9
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_001596_B_traj_001597/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4a37c5f27e80047ff649d0ec82602cb585262b3e59167c98cbaefcc73175a16
+size 11216138
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_002260_A_traj_002260/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_002260_A_traj_002260/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e47ee5f41e04b176da6845aca8f9301bc7f0d7b1
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_002260_A_traj_002260/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51c779020182933e13522486e8dccd211b5c63d51f9f05b1a56d7e4179fa556e
+size 7367208
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_002260_B_traj_002261/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_002260_B_traj_002261/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e11360bfa013852cba238eb2aeb5d338573357ee
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_002260_B_traj_002261/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b8c2401a711e58f76ea896c499ff58c4b6f83389c2cdf5d7b91908327b49a54
+size 8627040
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_003301_A_traj_003301/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_003301_A_traj_003301/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..35202a446e113670ba3d06eb2b308b115c33ef70
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_003301_A_traj_003301/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39c9e6a6737259b7a0da97363df307b004d4427e04d1cbd25b6575a92efd1816
+size 12492146
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_003301_B_traj_003302/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_003301_B_traj_003302/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cc21234873c20d291fba5cbc83c3c3779503d723
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260413_221314__pair_003301_B_traj_003302/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eba09d2c2b433fab17617263859e296634aaa1f9102d5ba46efb44787b399664
+size 11411345
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000154_A_traj_000154/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000154_A_traj_000154/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6c3b1290f2d9523010c72a2f317d040353d88e30
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000154_A_traj_000154/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ba2f691e77bca98e1defcd38179f6e1af4133fa5b0433bbbad01b97ab938eac
+size 16716587
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000154_B_traj_000155/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000154_B_traj_000155/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e5fa3486385476dc76ff6f7f2cd6a1edccb26546
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000154_B_traj_000155/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aba82d366fcfc4c721f173ab668a04b45d36e80f6b51194ce01a635f04b295cb
+size 25636419
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000172_A_traj_000172/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000172_A_traj_000172/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..7978720fd0b1c8759534cb0be7c21439b2bfede0
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000172_A_traj_000172/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90e8db9ce3aba699c5e092c657056adeaee6ec064e023939c323322d45d1993a
+size 9695763
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000172_B_traj_000173/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000172_B_traj_000173/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..01d07f823884532658f23042b472ad997f5072b4
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000172_B_traj_000173/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c0ca287ee24139e7ed54a90cc63c0eb6c9961a8746964c91c2cc8bee38461f0
+size 19030633
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000596_A_traj_000596/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000596_A_traj_000596/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..00d530fc77a782a2549643f265837c9580432eaf
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000596_A_traj_000596/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4a81df7143b9276181725c7dcf53d799a0d880a2c4c6405364f2df7c1001656
+size 26032332
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000596_B_traj_000597/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000596_B_traj_000597/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..57bc9438f232f894757ab429a21630a92cb8223b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_132732__pair_000596_B_traj_000597/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c450132b1efe5d83079b4aabea1b01c3d245fd74c20dbd54065571c523ed67de
+size 19024964
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_205421__pair_000934_A_traj_000934/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_205421__pair_000934_A_traj_000934/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..44c80a40a2641432d03c1b1f40b7c668fc78519a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_205421__pair_000934_A_traj_000934/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffa32f1dadb986d069ed6cd087e0d2c873270257bff649468ee13aa666dd66d3
+size 12816702
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_205421__pair_000934_B_traj_000935/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_205421__pair_000934_B_traj_000935/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9eae5bc4b6dd87678e5541a7a9a6f97a0f8dff10
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260414_205421__pair_000934_B_traj_000935/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3f9f7c3ea5fcf8a3166fa1d96f5093a68980094af3b45ca0cd265e7f20e4844
+size 13842012
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260415_212208__pair_000621_A_traj_000621/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260415_212208__pair_000621_A_traj_000621/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..5c8a177b4eed97634981a3d2af11350b1c65965a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260415_212208__pair_000621_A_traj_000621/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4148020e2f4a7a83671bff0329852e7ff62062c58dfd4ae388c07895efbdd50a
+size 23609500
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260415_212208__pair_000621_B_traj_000622/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260415_212208__pair_000621_B_traj_000622/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c5d262970fe6183c2446bbb7f36b92570726e96c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260415_212208__pair_000621_B_traj_000622/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d812affdcbf4f682d6a635e3a3a26709291053678bb9462c3ab9f97c3f5d2ff9
+size 29479919
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_001798_A_traj_001798/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_001798_A_traj_001798/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..bba6cf40880b9ad7504d767d7c006cb7e2cf97db
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_001798_A_traj_001798/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d4fbed4c0d617a11c3756647cb02c7f0f4d20d1e1f4c30cef96f3714cd152af
+size 10995061
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_001798_B_traj_001799/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_001798_B_traj_001799/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..f5266fa3abb1b561ee2f56b9d6a3680b7080d10a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_001798_B_traj_001799/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:271dd83e8669681d14a3193af3ae124b5e3825603a99b3e2b12e39c9800a30b3
+size 9915183
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002491_A_traj_002491/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002491_A_traj_002491/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..225d95c6ecea3e631aaf2a5001212caf7d791100
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002491_A_traj_002491/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c339fdbd4807521c238bcdf5b899b82f309c739529ca4f84da5efe3607f71359
+size 10176068
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002491_B_traj_002492/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002491_B_traj_002492/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1b30635d09098385900b9ff2b156d3b4a87f7cca
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002491_B_traj_002492/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d931afb366421ec7279625db3ff464edd3d36da95cf30367efc7cc728e9a3f44
+size 9012292
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002778_A_traj_002778/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002778_A_traj_002778/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..621c8ee47883c89e3f5814198d7370c1839cf27e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002778_A_traj_002778/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c673a7004f772379de80904fe9797c84758783c39b3009f778300dcb209260d
+size 9528725
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002778_B_traj_002779/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002778_B_traj_002779/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e9e7ab059ec9fc932a438f5771cf08ff763c9303
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002778_B_traj_002779/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:970f2948cfb5c8fdc2ba7631bf02d6321449bbc328f07c346043e768ec63ed15
+size 8241780
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002855_A_traj_002855/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002855_A_traj_002855/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..de4ce1d7d4687196fa207f977ddfa3762322f0e2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002855_A_traj_002855/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c598a80f25a0ddf26f8a1fba18d2a5a8da938142f9c2d06976220185fb76e984
+size 9646604
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002855_B_traj_002856/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002855_B_traj_002856/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c9b2d3cdb8ea2af9ecccc4fc8a3dff3b0c617e5f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_002855_B_traj_002856/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a52c33873305db08167aa2b9e18d5ffccd0c8babeea6b008f89c360d6ab070e3
+size 28261336
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_003390_A_traj_003390/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_003390_A_traj_003390/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cc5602de1f426e7a02237a9f14619c7d7642e770
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_003390_A_traj_003390/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8d0aac77a9e52804ba47084abe5548dd927286651b4a76ba0b993954aa436a0
+size 9426183
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_003390_B_traj_003391/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_003390_B_traj_003391/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..91465dbe89719709fe44a529dbaba5ebf93e62e7
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_003390_B_traj_003391/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8df8b8a256af8e5096afd704b9827a744b90e6c1885bcc5be79289b8f31bf9a3
+size 10979149
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004217_A_traj_004217/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004217_A_traj_004217/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c4b4526fd2b1317f8e8bc4ec1564d871422500d7
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004217_A_traj_004217/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:609277c74cbddd846b6f76ba2b40adeee6ca130efa624d71d60ae043d1d41d49
+size 7541511
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004217_B_traj_004218/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004217_B_traj_004218/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..02082956df4c03ad2b6fd9cea7a9438eb42f3ac0
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004217_B_traj_004218/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2e22259e349063c191981351d55dc462a1fcf9e763ca79a2c7a74d771e50833
+size 7056052
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004944_A_traj_004944/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004944_A_traj_004944/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..705fa1d549ab3494f716108c2e058b05ca8d5eb9
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004944_A_traj_004944/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10f2f8459da9301716e2114cea70bccba8cc4c013582a92bffb99709b432266e
+size 7303306
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004944_B_traj_004945/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004944_B_traj_004945/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b052800191c4e4a3245cfb646a557145eb394464
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_004944_B_traj_004945/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2e6c230dc6e4f0b8bbf97b1ce2c3306640a626d07a0125ab78fc8e328a9979f
+size 7388937
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005292_A_traj_005292/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005292_A_traj_005292/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6749f97bf317aca9a8cc21aa51a8d5c5ae260245
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005292_A_traj_005292/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edaf53c49f622f0ec1d903e677a13dc5f98c8a501b55462f82d7493a6416a986
+size 6548127
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005292_B_traj_005293/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005292_B_traj_005293/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0d7e8b0a1a97c0adcd8f2325fac408ca72e83a4a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005292_B_traj_005293/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22d74b8ed521495ffaa87693c1b66a436a37757f69f12f2b85e1d785a34b515e
+size 7514617
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005324_A_traj_005324/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005324_A_traj_005324/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c7977fe26593d43a7be4115b57be27f5da8474fe
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005324_A_traj_005324/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efde11b817f937f8087405e28eb5d682cdac30048200e9f64c3e242ff92d99df
+size 7263623
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005324_B_traj_005325/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005324_B_traj_005325/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..a5a97b21aa45e19e8f56de3fea1ff43ca3890e21
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/equivalence_easy/run_20260422_172644__pair_005324_B_traj_005325/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99c430545d58e01f52bb2d68182454086b76db6e3c02e6eb0add3aa622b94c44
+size 7409736
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260407_174321__traj_002355/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260407_174321__traj_002355/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..030e45c951fbaff9b2292592762488e6c8375615
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260407_174321__traj_002355/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46a34e6fb287eb5578e76f8aa5d5d7f59aae42a99dc6404ef018556a8570d178
+size 30988212
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260407_174321__traj_002454/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260407_174321__traj_002454/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..936cd05d56d0035376eab8e8992ba22ed38d4d2c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260407_174321__traj_002454/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90f4177228e20b11302653fe32ebbb8048009c76b9a6ae1d4a16d0badb0b8275
+size 22673165
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_001504/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_001504/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..53049ad0bd22e0f53513445c188406cfe6974f64
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_001504/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86507a32f6bf6220723afb9abe616559b9f19dd0ff3ad73ce5acd4f5afc23a08
+size 25573228
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_002567/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_002567/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..22dfc0bc7ae755005dc6e0cbb9a669a7ee1f558f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_002567/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:460b859106342302770d1743c35370d4f9f4edccf77bbbe8232a1ce44a7f8113
+size 18190589
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_003021/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_003021/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..68c8c7e4a9317ed60a3ef5ed38301b84ac6923b6
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_003021/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85616cd5a23b75c3299264adff4d10da930c9335ec09850883ddf43285b73fe7
+size 30027230
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_004105/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_004105/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..84a771d73dc04193a4ec676e4163898339f5925d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_004105/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b23d8860a965bef1e7950bf12dc12375ee4d8fc04a1ed42edb19598879234e4b
+size 18751393
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_005386/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_005386/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1419aa088d3b1e0b001b998c8b98047acac82f1e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260408_180248__traj_005386/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bf926d82ed3b9906c02980c661d175bc32d9af98cf9bb453e981959469d188b
+size 28056357
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260409_134704__traj_000431/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260409_134704__traj_000431/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d207ed0cebc4066b7fd82f7e19b8121c9a460d2f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260409_134704__traj_000431/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:715adb6d5e2d387b252f4d17f1030cddc71d7934d763a92ea37c8c4bfdae15cd
+size 15550781
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_001167/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_001167/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..038fd0c8224b6cbc3baea654f2eaa9545ad1f0de
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_001167/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ffca00864dd17d2b9f229d80866fdee61a4a43a81479aad170cb4f7aed0a75b
+size 36723284
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_001498/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_001498/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e66cdbcf86b8cd57f658f193f13e61bd687d9e6a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_001498/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b6a3ae923761ba98c9fd911afb24af227bce0327e4c9ac89577fa9264c55d18
+size 29337192
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_003759/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_003759/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..88ee5c2b86e67771420fe7a014f09f8dc938ffcd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_003759/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:598434a6f7a1f9574cec7b05d268894088731b7471ea000312ebcf9582b2beca
+size 23251728
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_003942/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_003942/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0311d331822986d5c6d61a956c9034e4a15b94b6
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_003942/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1386ea55ee2350bb58d1a9132377622d14e942444b1ce2ae840b54e3b6e9a0d4
+size 18065721
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_004594/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_004594/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9e99608be0f28a78d56327b3bcf7220d08d80f3b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_184831__traj_004594/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:503e9d830a4b048542adfc5167d171613526c08ab34082539a3405d4a55796ff
+size 6639483
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_204603__traj_000013/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_204603__traj_000013/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b2f02181182a69393648fdcb27265410f6cd9c4f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260410_204603__traj_000013/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:836f0a0e9ba3c70cb81672cacce8a64f24563cf56c5f9946672d9ea9a8f86424
+size 16116186
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260411_181558__traj_001520/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260411_181558__traj_001520/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..f181fc00fda18bf6122ef70e1411ad1386b34c23
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260411_181558__traj_001520/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ca56696fd870c9bd352eac3da481b1d60ab854be28575c3447492006b47d854
+size 30653638
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260411_181558__traj_003260/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260411_181558__traj_003260/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..5bbec6cf4b3e264a6bd44d03afb06eff32d99008
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260411_181558__traj_003260/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:982a1cefceb7549d3dd0abd95bdc8be0742b3a63f0bad922112dd19082b0f014
+size 26644025
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000766/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000766/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..2d27c53931cb2da004591f160abed570c46bf578
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000766/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed2ecaeba83088c3e3c4cf3cc3ca6189c5e3792f97134d6614fa2db750002bba
+size 30188294
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000826/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000826/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c8b2299135dedab472adf81d0edd748df440df21
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000826/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a080e9c0f9ff6e53826fabb252323f0cb760c7c190513ee01e3fa8685a9ea966
+size 5166307
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000839/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000839/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d577b69a0bf89d5aae4356f03cdc279a449f8b3e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_000839/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3cb163768525d7ffca3827931ee98264f62d776b8acbc2ae58de6e90a9527f7
+size 5798915
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_002041/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_002041/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..374702e9eeb287c955b2e8cc708f5bc4b3f718bd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260413_221314__traj_002041/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0407efc8369477df265f319ee012fab5a9ae1396cadd4ee2da7050b1a3b32b08
+size 19960107
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_132732__traj_000031/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_132732__traj_000031/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0dbdd284cc2dc14d05076a3bf8a069204033e7cd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_132732__traj_000031/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:203f30f15b371c7adca8703fbfcc78d7daa70b80a1f45c16c6450f1727346487
+size 8619026
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_205421__traj_000479/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_205421__traj_000479/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c33891f866abb647756d8d8a3191ecf10b48315d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_205421__traj_000479/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:697368f6be5050ed66fed154884efda031e3f8aff721cd981e2b59b9bec9c408
+size 25813055
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_205421__traj_001251/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_205421__traj_001251/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6e6441ee6bf1c2306a877099db307615a7520d7e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260414_205421__traj_001251/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47ffd6029ccff6ebb104b27cb2d3f9d412f1eae00020f2e1b6fbb313ce07e99c
+size 18567010
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_000024/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_000024/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ed4fade35a47ec568806db18363dc4304f27c3af
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_000024/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46a52243aef13b6535428eabeb382c4669757702168ce1abd3c44a88f435eebc
+size 6407861
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_000850/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_000850/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cf1386e6e42860d71ff1f15641e1199b9b913ba2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_000850/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed78dc177569ac1e31bda7047a777340126414b0396a069474948567837fe9af
+size 5561081
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_002919/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_002919/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c4910c9f52c5b6a7a2148aad046a8a1397f6ec07
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_002919/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3250e91b58b34ca5469e1a1073768dd8b218b2bb06c17ea255a45eb57d293a5d
+size 8370633
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_005311/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_005311/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..81a7ebaf436e24ee24a57bf6e950b62963d743ea
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_005311/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d74c17db54702f22af5e7c9275e74a24089a4f6169f110d6bf35e37ec0632f21
+size 8865486
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_006125/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_006125/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..00469370fb6812694e9fa3165b52a888dfe5664f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_easy/run_20260422_172644__traj_006125/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bfa7738b324fbfc9cdf24dd1b51b1bfab43560e85c0a3836d5581da1c4fe3e8
+size 5949491
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260426_170845__traj_000182/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260426_170845__traj_000182/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..fd6f20be19ad3eaef9436a40ea00cf8f51b82e57
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260426_170845__traj_000182/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9025e1e59b2c33ef7a816a53adbf36fcedf79b9f8e0a573468be417c297c5eb2
+size 25752235
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260426_170845__traj_000486/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260426_170845__traj_000486/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..669fb67c15c39c44d852de1b4d494f29bf7c757f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260426_170845__traj_000486/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc8121a2be785266c35433092a7c61bb9885d1004291b4fd6cb4d6e643819dd1
+size 21906676
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000029/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000029/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..fd374867f8a28955d50e40530184a118f9520be1
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000029/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dc0e1179b99c24997fb9b6163503026d3a9c8c9dc47d61e092453afd1068a59
+size 31231356
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000047/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000047/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1b1b438316f230aa6d9f6012a816b2df4bcef029
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000047/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eac1f324170e2c5ac92d235beadf2eceaf42993876554a2f51fc8f8ef40fd945
+size 5512714
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000130/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000130/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1bf2882066aa2624006b4be9fab39b9882306e58
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000130/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1d2252cc8d1829d92ebf0752738f75be8bd038526cfefeb57a5ee8081140a70
+size 29713971
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000183/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000183/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ba1cd597227272e24a10224d20757a45b34506c7
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000183/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:886e9bf3809412faac70585e0beff39addfcad42309dca8dad6054f21c51d1bc
+size 35968518
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000188/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000188/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..3befd60935af4191d475082578e66e7847b411f8
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000188/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d1b282181a4423b178c26610aa2075ceddb4d8c28f1b8091b0247002af713b2
+size 36535023
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000216/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000216/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..157527552e48e5747f33eb929b2e4cd5a0c53299
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000216/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d709b8bee97bf35d1ef3b72430ffb5ca97070b27ed04e8581c7f835c861053a
+size 22113307
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000255/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000255/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..bc25f2b31346d888bf2bcc0ed81164fe95c45ebe
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000255/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7d746017d8bbbc3679be47dbd349ea406eff990640ad3c8dbce380ad378f205
+size 10399803
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000303/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000303/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..3bb282ea36870cf1c8eea2932d0c8844730aeab3
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000303/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65a41548bd37c7078fa5d95b250b249710a65fc2707f63eeb0d9c6dbfea2962c
+size 31749129
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000305/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000305/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..daa059fc9a5105f11380edeb36731e0f5cc9d936
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000305/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbde962c97abbdf62555893a31e2072e4d398aa807b8d13b85488169e7dc35d5
+size 11614999
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000344/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000344/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..5423b494f38a49df40bb931a86eb7a0f051116e8
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000344/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6958fc8703c6692a6e4d283c8768ae8fe43713014a39de44c019a1b3b6a36edb
+size 23373016
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000349/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000349/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b373950fcc2e490d7ad8e617b1755877083a6865
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000349/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bea9beea3f345ff3c54c623c161db91797ec711dc4838ef1a5a539227f4e1314
+size 8391485
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000389/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000389/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0212f801d238c0809ec6cb238bd0297b8336df57
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000389/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23d83c3822782b4a5f602f8c992e9dd6d14faaaed80fab469a9accb37409c5b9
+size 19386016
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000391/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000391/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..152f6c587ddb97358b1c37c6bc9aef9956de4223
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000391/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b5ff4d7bb9428b679f967fd82b43d17a3dc4187048ff6e928605f954f18ff5
+size 41076287
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000397/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000397/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..042653146e9538ca68769cbb0bf380ef2a2b40ec
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000397/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9455eea365d9d9594c3e04238a81621e190b5c9925737945708593f7a815d00a
+size 33221749
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000419/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000419/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..37f89568a98c207a2727217f10267ff2abc10ed2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000419/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3b5bf404b104b8d06645d9cee17a60836e8e2859de6c9d0baeebdf45e5a5788
+size 6315563
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000462/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000462/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c5d42f341fcd93452074fe53f5f6eba06ae362b4
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000462/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:571e103f0e65d01808a2d56bdf8477c15d757e40e8e1c1c88cf5b3fed27210d1
+size 6840140
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000485/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000485/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..955b61486a14fa6fc292bef40a7523abcfb1e844
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000485/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36ae1686273783fd2c59c86080fad2cae602b22ae11ffbc48aa66525444277a7
+size 7955990
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000501/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000501/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..73a3acfcadca268dbf01a8acac006af22fd5ae93
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000501/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6be3528fedc92b20a207b1071185a05470724f161c8ce64b1fee034bd2c3e50d
+size 11321287
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000557/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000557/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4a6664dc974029e3d244d8103ffd2fbe269b1386
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000557/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f4abf8e1f963cb3a811b62c672d9f07e5d560d5efbc97a29dcc13634f3e4af6
+size 27124428
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000569/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000569/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..3a777c5239bd6599dd8f238350c80ed2a0e1ddc6
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/inverse_hard/run_20260427_001012__traj_000569/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4afbdbb6bdbfe5d02aafa0f82f00522dbc6ece0d72c63bc7d5b6624115f81111
+size 30410164
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_132406__traj_000436/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_132406__traj_000436/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..193b1824c394b1c104c7bd6c7f62c0508856df01
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_132406__traj_000436/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b8c8ae402fd7d4c450933fe575cb49cf57f708856afe8f30a76d8d8b8066575
+size 21294084
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_132406__traj_000496/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_132406__traj_000496/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..169cf370fdac87f57c08fc096c7b8690022eee6b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_132406__traj_000496/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32c9a8d092a85eb80115af200e1ad9b6a8f7001fe4ab757da04cad0c0445f539
+size 40701024
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_174321__traj_003040/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_174321__traj_003040/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0e131de2a4c184d8f57353f06a36a06b17486ee3
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260407_174321__traj_003040/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e3251f6a1cba31ad4d44583a1bebe63998823ddba4cfa7322008fa2b63e662a
+size 21001705
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_000744__traj_000691/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_000744__traj_000691/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d5ca9073899cb4b8838f90be13f99da6ded8fec6
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_000744__traj_000691/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cd311254228265a415aebccd18938999a0560f4540607e6da6b95d729660b61
+size 29959062
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_000610/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_000610/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..068b5a47bcc6a9393b948a62262b6d0d51c82bce
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_000610/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98883677ed66e4cca9f72a130d029345b166aeb88e69d9f7c0d44cb4773203b5
+size 23816651
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_002353/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_002353/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1f9f22f262214ab34e60c99375936676cd650825
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_002353/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c1fd25128715dea349e4e1c2e071bef5bcf87b88c0e9476b81625dccd5bdf40
+size 4140666
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_004617/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_004617/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..be90069cdf335f17d6ff0b0bf178630262b78ccf
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_004617/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:042ec5685ca5cd75f418f4e2a687335f7ca4f82380cc41a6140f812408b16a95
+size 25981018
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_005859/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_005859/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..3aef43f4108e3ac6d32646319e7a07b7249a6e2e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260408_180248__traj_005859/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b2d465b0a2983f14236f49a62a4c17902bea21f769b6c62507cc8ca0658ec6e
+size 29818409
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260409_134704__traj_000435/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260409_134704__traj_000435/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ee6d0234bb897e21ea6c3007506317f415780652
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260409_134704__traj_000435/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8583dc24785aed6b251ea6a717d91d93951778f5a7d93c5d0d53b0ac03466c76
+size 14289015
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000147/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000147/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6a00630c2084cacb2dbc5ccb82f44283f921d8b8
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000147/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d6de0530ef9373eb5d7e0065667050e7fa1e40ffb1d975bc3f8774d5f1c0d5f
+size 11719592
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000272/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000272/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..be7c7bc0718bec4e49dce88440026a8c3f6972fe
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000272/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70b3b08407402f31c189f1d8724bb4cc0005f5ea68856429965f6764e0d5e4c1
+size 14161595
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000274/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000274/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ea1e871254ac18bbdcd7ddd26f03552945800c60
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000274/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58ce7623bf54b8aadcbd6155629211012a77084c1284f54fb1c24e6df4829013
+size 21121988
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000454/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000454/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9687ba344d1d6bc73e3a2a0cbe42d14754961734
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_112607__traj_000454/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3143a1ef4b66f392a09b2569a1e348a88a94b69a1dd06415a7b9d72aa87dca63
+size 11979476
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001329/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001329/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e2471251c0c9d82cfe40842d466968731305ac5d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001329/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fa33c223ffa8422d610b27304c00c4e45932ed16e6ffe53a6fb9c3bcf5e3f57
+size 4326874
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001649/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001649/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..422a1ec9134a5557c639800eb3715f0a399abd2c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001649/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31cf5f82285ed508fc8fc131e7d92ae6616bc25ab1ed73421e601c5eadd952cd
+size 15190345
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001671/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001671/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..371ea81f899b20a6a22d154595903814725a9415
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001671/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:515893ebab1bdce7157eaf1ea04789188c690f782de77a6cce8a5b5cb7c4fa7d
+size 7096966
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001997/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001997/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b1c5c67212e9bebf441a6d1eef312f7b40d3749c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_001997/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a40dcd29e2e26586d2211d52a1e6b91ddc894f6e67a0994c832ae2aa94ea70ab
+size 3401750
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_003329/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_003329/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..7c9eb4075f26c7ca1a4d148f2fcee133d3aa3b3d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_003329/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:564612168331756a5fed0e6e81a97e17036c8d663f80c18729b311b72ee2598e
+size 29857761
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_003388/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_003388/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d036bb6220aaee51f1dafe44c8fe1b375bc382c4
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260410_184831__traj_003388/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d1ff2db241beac59045bd36be7b7cc870f3cb3a3f3e7666f95161bb06a021ae
+size 3575199
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260413_221314__traj_001051/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260413_221314__traj_001051/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b442ff0be025117da234e0fc8dc9ac549ea525ad
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260413_221314__traj_001051/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57e353afdeaf5982fb1683f8e50de073354223f8f464ad6b627df115f6d81ed6
+size 3204642
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260413_221314__traj_002379/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260413_221314__traj_002379/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..eb4103f8bcab4f26242914fead7728b54f2d1be1
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260413_221314__traj_002379/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:356051df546158318a7d6c928347eb2786ac4ee8ff895b7be9dc4dfffc768178
+size 13073627
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260415_212208__traj_001809/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260415_212208__traj_001809/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..fb65fed0444c52513a20017b95617eeb8f5851d7
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260415_212208__traj_001809/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c7350ce6687630d6a2bd3474121945880123e13187da7b1dc45b2585e431baf
+size 35061623
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260419_133642__traj_000462/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260419_133642__traj_000462/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..56ba384156b7cdf3740998f8e28b4116d23f0ca5
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260419_133642__traj_000462/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40ba4466e185bbaebe26b80b3fde4a9013d1d00a8790085b43eabb937f0c77cd
+size 4544512
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_002006/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_002006/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..7bdcfed68b30c0ee7a968b6b0c5712bf0dc26abb
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_002006/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34559eaa3da1cc1551c05787901e09b39c244cfbaf67cabb1f5f30e1998ae318
+size 7172238
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_002734/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_002734/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..15bace22cdca6f61975b97d2883bdf94dc367041
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_002734/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39a9d7d6515aca63fffe64126c8bbc1af351fda56d956d0c6a059a59c993ac6a
+size 7994445
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_003068/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_003068/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9481754b697414142bbddac83dfbe9b7daeb3b71
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_003068/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b589d2e609a030fef0d4e8e26368cc38658af66095f8f51c87f65f617d7bff6
+size 9070940
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_003591/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_003591/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d1b928206476c9b34d4f8d45c79a313af49b91e4
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_003591/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83662d2e298953ff1c1142b642f90d33b9a0791824a4fc8425c13c2deef370ad
+size 9650059
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_006296/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_006296/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..12adf76b51a31463cfda7d316eba113cbedbdf76
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_easy/run_20260422_172644__traj_006296/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cfcde950a549e9a7c01da20b61e95aafe5799cf23b00a522619034e257476b9
+size 5705353
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000013/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000013/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..ff26f659cfed16a7d0b3b283a1a695a27574d318
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000013/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2aa6a3ff441e1f5633592ecf27190aeb2154a140e688f5a3c10be15e72afb7ed
+size 40044556
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000037/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000037/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..799106a9f0753267120f8121e31074637c595840
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000037/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5ac1a292347cb254086cff8889e29a04115afea7ac86aca50b2e59b67cdfb90
+size 39363588
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000311/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000311/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..a53abd9eb3e5f910b87a2f1c4e23e7188d857532
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000311/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43ffb007069d4f1e7f56b8481e5276cd73bb33ecbae3b7585b2a2e0648c27510
+size 34629262
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000388/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000388/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9b93c2a554bb544988e6ee03ac439886dadcb511
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000388/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6140a4367e639135abdf5124fed6484f9755c7588e484ea4776e5beb6786c0bf
+size 14396390
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000397/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000397/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..7ce922454cd500440ef6fee63ad27572fd430d3a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000397/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc7a51cefc2dc42ee8de7c94b548a4e203336da4c80ed0cb1648875838b89d38
+size 22814472
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000408/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000408/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..602b9abb2f2b0a6373345f674a0a61d1e0f96bf9
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000408/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff2405158a42973243630dabde4cca826bf11fabee5582fd82f0dbb127da7bf7
+size 29392536
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000418/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000418/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..f0ad78f35472a6c1818ffedbb909050f057abe86
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000418/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b6d663a0bb20508eb99f7e89b4c75d7b4b1d958d1f8ccb5d14ae10853a51828
+size 18600431
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000468/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000468/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..2113a9983e4e34faf17c59b281669524dd09810f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000468/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23417f6b7cafa42cc8357d4f771383cb651c5ba687a2ddcf86d5254e9494911f
+size 23943259
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000506/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000506/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..afe060adc718acf6edf1adb06201b63fa0f84b4d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260426_170845__traj_000506/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0795ea509805b4922b0bfe62f2d6cc6927caca0a4bf9c2045fad4923e1241bc
+size 23144245
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000068/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000068/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..57bd013cc72208c873d3c28e77cbfa4b1640ad08
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000068/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f59ff424da609d4982e1367ef6656e2f3025b0dccc925edb127b1d84bcc330b0
+size 7925967
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000133/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000133/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4ce4e13daf123ff2bb7f70bd2deba4a094e085cf
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000133/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5725c07f0af52ab7dc5d4cd17696a505cfaa7a4656730f5b0fe6375e5dfdc82
+size 15852560
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000138/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000138/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..104f6085db193c7246a9cc4fd23df43f54094db8
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000138/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cafe9a88a94e1737c4326686bf27179381222bc607d0ecaac88f5c59dbbf67c
+size 10262215
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000161/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000161/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..77b4d36258d6f714ed88551194710cd75eb2f5de
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000161/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d17254a0b8291f6dd5b52424730a89a23c42076fef6eed5243cde2e8457168f2
+size 32940571
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000162/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000162/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..256a96c3d64db6ad8ba098878b926fe7000a54ad
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000162/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fedaeed93f108a1dc43f23b4a86568474ffacc3a5313a381e07e80edf39ccac
+size 26842666
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000194/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000194/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4d44e5eeb39c220c106e6f1679411d87f8e3febd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000194/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e13202c4e3762e74c69d038f0610ee1a2db66d383432c7987c54aac12f84bb5b
+size 26304131
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000225/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000225/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..acc1863f1b9e78e1f701e91061850134615d0f34
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000225/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33ceb2409afd4fb6ecfdb4cbd309d1ecbee354c14520fff903540396c7c72402
+size 32175516
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000259/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000259/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c2e62655d728c08f6581c760c9a4143ae8dd8e01
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000259/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:079d7b8459221dad3431987c568e3f400c0f845e93843a917359433567f8fccd
+size 12843610
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000271/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000271/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..42169797068ac1aa9fdd293cdf364e1bba47fbbf
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000271/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e80635367944de53c692ab1fe7ff574a21f649d4c91739884deed2035439ca2e
+size 14020848
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000335/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000335/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..39c6e182f762560b713f9e38f62954b54d5b51dc
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000335/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bc78804991e7a74b94f316287469c42dbbf5fd07ea2d901f5f0f9890bdabcb3
+size 9607365
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000357/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000357/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9c5cf78ad123916f7242a7a1dcf5eb168bdaa98d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000357/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd5686971e2004791a1f03b6575fac13d5cc187949cdc9190817448cfcf3ff44
+size 9522377
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000451/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000451/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..7c6304089e4286bd809e6182ab2b0e4269995a06
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000451/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc67f65eb099e12220df7b6a0b91cfad8d623faa959807200f292cb75fd3af13
+size 10723208
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000494/rollout.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000494/rollout.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..56b28928176dd48346eb4561bea50b3d0b8c0dfe
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/predictions/loop_hard/run_20260427_001012__traj_000494/rollout.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbb5966886574cd325600a05a420cf290d1e32e64d37aec607e1b1a511a9a75f
+size 9641496
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260407_174321__pair_001790_A_traj_001790.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260407_174321__pair_001790_A_traj_001790.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4775db80bcafa6c64899bdb31d3d573b4666bf41
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260407_174321__pair_001790_A_traj_001790.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:044b3f3424622934817b310ce52a8266ac13eb36148cd43e1ca02444fd72e8e1
+size 15510493
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260407_174321__pair_001790_B_traj_001791.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260407_174321__pair_001790_B_traj_001791.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..49c5ab13cb3c1e5046c058a76f5b4fec500b9fdd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260407_174321__pair_001790_B_traj_001791.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77dd773a265468e1f4a779b6867595ca6848cb37fb13710d7fe7a4aa66c07c32
+size 15839627
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_000744__pair_000476_A_traj_000476.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_000744__pair_000476_A_traj_000476.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..404167e1b268a319fda11a32fb3095b7223b41dd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_000744__pair_000476_A_traj_000476.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5921555762ad44dbf363307d7516e7d9e55eda90388c2f199ef89726fc04616
+size 34672950
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_000744__pair_000476_B_traj_000477.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_000744__pair_000476_B_traj_000477.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..86a623c07a92dac46437a45101518f90ae9c7212
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_000744__pair_000476_B_traj_000477.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfc6c017bc86f9cd31f449f15e54c2a40fae13c7f01fcd22589804e014b9c2be
+size 34953965
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_002128_A_traj_002128.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_002128_A_traj_002128.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cc8b5f72611230f526e3e19963f3902e43d01683
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_002128_A_traj_002128.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2a8e338ff7683b625605516ec744f4acf8cd43902d5670dae5f2e5a84e951ac
+size 17982802
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_002128_B_traj_002129.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_002128_B_traj_002129.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..a057ea4bcad36e18d3b4eb6d5b1a30c084780b8d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_002128_B_traj_002129.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bb61ec65132ab130737e6db9ff045e76764180586edea3bc89fec0f995e362c
+size 28815412
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_004149_B_traj_004150.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_004149_B_traj_004150.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..86958e11b9f9ea273548466b2d7c8f70ada7e22b
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260408_180248__pair_004149_B_traj_004150.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9734917c24cd9a68d407828d8936faa28ded709eb52d3e56608fc0d5022245a5
+size 26421539
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260410_112607__pair_000166_B_traj_000167.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260410_112607__pair_000166_B_traj_000167.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..00338ebba687639a579614968e1a4ae752d19512
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260410_112607__pair_000166_B_traj_000167.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c4ccc1a1dd305a7fccf03c4751cd24775287f73e5364b6ec4f03e775ddd552b
+size 19992000
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260410_184831__pair_005031_A_traj_005031.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260410_184831__pair_005031_A_traj_005031.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..22f80be8733b86acd9ac4c928f5d52f9ab305166
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260410_184831__pair_005031_A_traj_005031.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:774d02cdb9bce3792201c160645c60611e49bd2326ff090a00f88db2ad1e2807
+size 32258521
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001130_B_traj_001131.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001130_B_traj_001131.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..129c46ff4fd28995b9b5ad203a5bef80ad421116
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001130_B_traj_001131.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:348f557a5e56f6bc39334e471019ed4e28ade0334799fe14bd42d1013228cfe6
+size 20199497
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001156_A_traj_001156.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001156_A_traj_001156.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8e40f592890cf77ca5ed349aff70295f029ae139
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001156_A_traj_001156.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a285324b4711fa3d2064a05f1cbe49c589e562f302a742e1d07d8bb93ebc6848
+size 18260347
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001156_B_traj_001157.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001156_B_traj_001157.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b95f0a2d8e8ab8ad1924d70797959d89164634c5
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_001156_B_traj_001157.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94bdd73c7273bc8641f997e7841742180cd68883abf53d9a526d9f34c2c8b18b
+size 17609839
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_002170_A_traj_002170.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_002170_A_traj_002170.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..10f1d721b573a77f720e9a6775af66ec70e06070
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_002170_A_traj_002170.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6cc58eaa2acdd1f1b4577360698eba9ab14540d0aec2142fc4cfa50835d057c
+size 18111479
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_002170_B_traj_002171.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_002170_B_traj_002171.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..8a515bfe988b2f0aeaa3c6dbc17a395ff9f72021
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260411_181558__pair_002170_B_traj_002171.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc3c7cf19ae30d6f4b4641bd9e3d38d72274f8b19991af69f3130455ae529fc6
+size 16348978
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_001596_A_traj_001596.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_001596_A_traj_001596.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..4af514c076594716d5943776996ecef648aa9520
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_001596_A_traj_001596.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8c7a91709917bfc9cac25fd2591d681307a5bd45ab58026047b68cb1f201160
+size 9321128
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_001596_B_traj_001597.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_001596_B_traj_001597.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..786fc866a22a197de9279aa150eda7a3f57725d9
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_001596_B_traj_001597.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4a37c5f27e80047ff649d0ec82602cb585262b3e59167c98cbaefcc73175a16
+size 11216138
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_003301_A_traj_003301.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_003301_A_traj_003301.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..35202a446e113670ba3d06eb2b308b115c33ef70
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_003301_A_traj_003301.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39c9e6a6737259b7a0da97363df307b004d4427e04d1cbd25b6575a92efd1816
+size 12492146
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_003301_B_traj_003302.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_003301_B_traj_003302.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cc21234873c20d291fba5cbc83c3c3779503d723
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260413_221314__pair_003301_B_traj_003302.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eba09d2c2b433fab17617263859e296634aaa1f9102d5ba46efb44787b399664
+size 11411345
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260414_205421__pair_000934_A_traj_000934.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260414_205421__pair_000934_A_traj_000934.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..44c80a40a2641432d03c1b1f40b7c668fc78519a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260414_205421__pair_000934_A_traj_000934.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffa32f1dadb986d069ed6cd087e0d2c873270257bff649468ee13aa666dd66d3
+size 12816702
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260414_205421__pair_000934_B_traj_000935.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260414_205421__pair_000934_B_traj_000935.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..9eae5bc4b6dd87678e5541a7a9a6f97a0f8dff10
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260414_205421__pair_000934_B_traj_000935.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3f9f7c3ea5fcf8a3166fa1d96f5093a68980094af3b45ca0cd265e7f20e4844
+size 13842012
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260415_212208__pair_000621_A_traj_000621.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260415_212208__pair_000621_A_traj_000621.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..5c8a177b4eed97634981a3d2af11350b1c65965a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260415_212208__pair_000621_A_traj_000621.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4148020e2f4a7a83671bff0329852e7ff62062c58dfd4ae388c07895efbdd50a
+size 23609500
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260415_212208__pair_000621_B_traj_000622.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260415_212208__pair_000621_B_traj_000622.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c5d262970fe6183c2446bbb7f36b92570726e96c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260415_212208__pair_000621_B_traj_000622.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d812affdcbf4f682d6a635e3a3a26709291053678bb9462c3ab9f97c3f5d2ff9
+size 29479919
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002491_A_traj_002491.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002491_A_traj_002491.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..225d95c6ecea3e631aaf2a5001212caf7d791100
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002491_A_traj_002491.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c339fdbd4807521c238bcdf5b899b82f309c739529ca4f84da5efe3607f71359
+size 10176068
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002491_B_traj_002492.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002491_B_traj_002492.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..1b30635d09098385900b9ff2b156d3b4a87f7cca
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002491_B_traj_002492.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d931afb366421ec7279625db3ff464edd3d36da95cf30367efc7cc728e9a3f44
+size 9012292
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002778_A_traj_002778.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002778_A_traj_002778.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..621c8ee47883c89e3f5814198d7370c1839cf27e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002778_A_traj_002778.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c673a7004f772379de80904fe9797c84758783c39b3009f778300dcb209260d
+size 9528725
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002778_B_traj_002779.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002778_B_traj_002779.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e9e7ab059ec9fc932a438f5771cf08ff763c9303
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_002778_B_traj_002779.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:970f2948cfb5c8fdc2ba7631bf02d6321449bbc328f07c346043e768ec63ed15
+size 8241780
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004217_A_traj_004217.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004217_A_traj_004217.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c4b4526fd2b1317f8e8bc4ec1564d871422500d7
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004217_A_traj_004217.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:609277c74cbddd846b6f76ba2b40adeee6ca130efa624d71d60ae043d1d41d49
+size 7541511
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004944_A_traj_004944.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004944_A_traj_004944.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..705fa1d549ab3494f716108c2e058b05ca8d5eb9
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004944_A_traj_004944.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10f2f8459da9301716e2114cea70bccba8cc4c013582a92bffb99709b432266e
+size 7303306
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004944_B_traj_004945.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004944_B_traj_004945.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b052800191c4e4a3245cfb646a557145eb394464
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_004944_B_traj_004945.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2e6c230dc6e4f0b8bbf97b1ce2c3306640a626d07a0125ab78fc8e328a9979f
+size 7388937
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005292_A_traj_005292.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005292_A_traj_005292.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6749f97bf317aca9a8cc21aa51a8d5c5ae260245
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005292_A_traj_005292.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edaf53c49f622f0ec1d903e677a13dc5f98c8a501b55462f82d7493a6416a986
+size 6548127
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005292_B_traj_005293.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005292_B_traj_005293.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..0d7e8b0a1a97c0adcd8f2325fac408ca72e83a4a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005292_B_traj_005293.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22d74b8ed521495ffaa87693c1b66a436a37757f69f12f2b85e1d785a34b515e
+size 7514617
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005324_A_traj_005324.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005324_A_traj_005324.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c7977fe26593d43a7be4115b57be27f5da8474fe
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005324_A_traj_005324.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efde11b817f937f8087405e28eb5d682cdac30048200e9f64c3e242ff92d99df
+size 7263623
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005324_B_traj_005325.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005324_B_traj_005325.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..a5a97b21aa45e19e8f56de3fea1ff43ca3890e21
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/equivalence_easy__run_20260422_172644__pair_005324_B_traj_005325.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99c430545d58e01f52bb2d68182454086b76db6e3c02e6eb0add3aa622b94c44
+size 7409736
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260407_174321__traj_002355.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260407_174321__traj_002355.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..030e45c951fbaff9b2292592762488e6c8375615
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260407_174321__traj_002355.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46a34e6fb287eb5578e76f8aa5d5d7f59aae42a99dc6404ef018556a8570d178
+size 30988212
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260408_180248__traj_001504.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260408_180248__traj_001504.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..53049ad0bd22e0f53513445c188406cfe6974f64
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260408_180248__traj_001504.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86507a32f6bf6220723afb9abe616559b9f19dd0ff3ad73ce5acd4f5afc23a08
+size 25573228
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260408_180248__traj_004105.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260408_180248__traj_004105.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..84a771d73dc04193a4ec676e4163898339f5925d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260408_180248__traj_004105.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b23d8860a965bef1e7950bf12dc12375ee4d8fc04a1ed42edb19598879234e4b
+size 18751393
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260409_134704__traj_000431.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260409_134704__traj_000431.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d207ed0cebc4066b7fd82f7e19b8121c9a460d2f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260409_134704__traj_000431.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:715adb6d5e2d387b252f4d17f1030cddc71d7934d763a92ea37c8c4bfdae15cd
+size 15550781
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_184831__traj_001498.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_184831__traj_001498.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..e66cdbcf86b8cd57f658f193f13e61bd687d9e6a
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_184831__traj_001498.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b6a3ae923761ba98c9fd911afb24af227bce0327e4c9ac89577fa9264c55d18
+size 29337192
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_184831__traj_003759.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_184831__traj_003759.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..88ee5c2b86e67771420fe7a014f09f8dc938ffcd
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_184831__traj_003759.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:598434a6f7a1f9574cec7b05d268894088731b7471ea000312ebcf9582b2beca
+size 23251728
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_204603__traj_000013.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_204603__traj_000013.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..b2f02181182a69393648fdcb27265410f6cd9c4f
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260410_204603__traj_000013.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:836f0a0e9ba3c70cb81672cacce8a64f24563cf56c5f9946672d9ea9a8f86424
+size 16116186
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260411_181558__traj_001520.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260411_181558__traj_001520.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..f181fc00fda18bf6122ef70e1411ad1386b34c23
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260411_181558__traj_001520.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ca56696fd870c9bd352eac3da481b1d60ab854be28575c3447492006b47d854
+size 30653638
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260413_221314__traj_000839.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260413_221314__traj_000839.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..d577b69a0bf89d5aae4356f03cdc279a449f8b3e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260413_221314__traj_000839.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3cb163768525d7ffca3827931ee98264f62d776b8acbc2ae58de6e90a9527f7
+size 5798915
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260414_205421__traj_000479.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260414_205421__traj_000479.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..c33891f866abb647756d8d8a3191ecf10b48315d
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260414_205421__traj_000479.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:697368f6be5050ed66fed154884efda031e3f8aff721cd981e2b59b9bec9c408
+size 25813055
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260414_205421__traj_001251.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260414_205421__traj_001251.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..6e6441ee6bf1c2306a877099db307615a7520d7e
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260414_205421__traj_001251.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47ffd6029ccff6ebb104b27cb2d3f9d412f1eae00020f2e1b6fbb313ce07e99c
+size 18567010
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260422_172644__traj_000850.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260422_172644__traj_000850.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..cf1386e6e42860d71ff1f15641e1199b9b913ba2
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_easy__run_20260422_172644__traj_000850.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed78dc177569ac1e31bda7047a777340126414b0396a069474948567837fe9af
+size 5561081
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_hard__run_20260427_001012__traj_000255.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_hard__run_20260427_001012__traj_000255.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..bc25f2b31346d888bf2bcc0ed81164fe95c45ebe
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/inverse_hard__run_20260427_001012__traj_000255.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7d746017d8bbbc3679be47dbd349ea406eff990640ad3c8dbce380ad378f205
+size 10399803
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/loop_easy__run_20260410_184831__traj_001649.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/loop_easy__run_20260410_184831__traj_001649.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..422a1ec9134a5557c639800eb3715f0a399abd2c
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/loop_easy__run_20260410_184831__traj_001649.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31cf5f82285ed508fc8fc131e7d92ae6616bc25ab1ed73421e601c5eadd952cd
+size 15190345
diff --git a/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/loop_hard__run_20260426_170845__traj_000311.mp4 b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/loop_hard__run_20260426_170845__traj_000311.mp4
new file mode 100644
index 0000000000000000000000000000000000000000..a53abd9eb3e5f910b87a2f1c4e23e7188d857532
--- /dev/null
+++ b/inference/2026-09-23/outputs/miniworld_1b_r50_360p_v2/sc/raw/pred/loop_hard__run_20260426_170845__traj_000311.mp4
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43ffb007069d4f1e7f56b8481e5276cd73bb33ecbae3b7585b2a2e0648c27510
+size 34629262
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_002128_B_traj_002129/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_002128_B_traj_002129/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_002128_B_traj_002129/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004149_A_traj_004149/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004149_A_traj_004149/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004149_A_traj_004149/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004149_B_traj_004150/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004149_B_traj_004150/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004149_B_traj_004150/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004865_A_traj_004865/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004865_A_traj_004865/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004865_A_traj_004865/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004865_B_traj_004866/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004865_B_traj_004866/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_004865_B_traj_004866/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_005916_A_traj_005916/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_005916_A_traj_005916/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_005916_A_traj_005916/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_005916_B_traj_005917/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_005916_B_traj_005917/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_005916_B_traj_005917/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_006244_A_traj_006244/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_006244_A_traj_006244/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_006244_A_traj_006244/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_006244_B_traj_006245/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_006244_B_traj_006245/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260408_180248__pair_006244_B_traj_006245/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000315_A_traj_000315/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000315_A_traj_000315/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000315_A_traj_000315/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000315_B_traj_000316/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000315_B_traj_000316/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000315_B_traj_000316/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000657_A_traj_000657/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000657_A_traj_000657/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000657_A_traj_000657/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000657_B_traj_000658/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000657_B_traj_000658/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000657_B_traj_000658/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000778_A_traj_000778/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000778_A_traj_000778/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000778_A_traj_000778/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000778_B_traj_000779/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000778_B_traj_000779/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260409_134704__pair_000778_B_traj_000779/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000166_A_traj_000166/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000166_A_traj_000166/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000166_A_traj_000166/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000166_B_traj_000167/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000166_B_traj_000167/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000166_B_traj_000167/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000210_A_traj_000210/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000210_A_traj_000210/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000210_A_traj_000210/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000210_B_traj_000211/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000210_B_traj_000211/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000210_B_traj_000211/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000368_A_traj_000368/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000368_A_traj_000368/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000368_A_traj_000368/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000368_B_traj_000369/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000368_B_traj_000369/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_112607__pair_000368_B_traj_000369/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_002375_A_traj_002375/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_002375_A_traj_002375/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_002375_A_traj_002375/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_002375_B_traj_002376/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_002375_B_traj_002376/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_002375_B_traj_002376/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003641_A_traj_003641/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003641_A_traj_003641/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003641_A_traj_003641/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003641_B_traj_003642/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003641_B_traj_003642/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003641_B_traj_003642/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003864_A_traj_003864/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003864_A_traj_003864/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003864_A_traj_003864/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003864_B_traj_003865/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003864_B_traj_003865/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_003864_B_traj_003865/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004238_A_traj_004238/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004238_A_traj_004238/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004238_A_traj_004238/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004238_B_traj_004239/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004238_B_traj_004239/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004238_B_traj_004239/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004544_A_traj_004544/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004544_A_traj_004544/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004544_A_traj_004544/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004544_B_traj_004545/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004544_B_traj_004545/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_004544_B_traj_004545/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_005031_A_traj_005031/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_005031_A_traj_005031/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_005031_A_traj_005031/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_005031_B_traj_005032/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_005031_B_traj_005032/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260410_184831__pair_005031_B_traj_005032/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000020_A_traj_000020/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000020_A_traj_000020/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000020_A_traj_000020/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000020_B_traj_000021/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000020_B_traj_000021/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000020_B_traj_000021/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000896_A_traj_000896/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000896_A_traj_000896/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000896_A_traj_000896/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000896_B_traj_000897/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000896_B_traj_000897/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000896_B_traj_000897/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000975_A_traj_000975/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000975_A_traj_000975/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000975_A_traj_000975/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000975_B_traj_000976/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000975_B_traj_000976/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_000975_B_traj_000976/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001130_A_traj_001130/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001130_A_traj_001130/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001130_A_traj_001130/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001130_B_traj_001131/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001130_B_traj_001131/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001130_B_traj_001131/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001156_A_traj_001156/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001156_A_traj_001156/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001156_A_traj_001156/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001156_B_traj_001157/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001156_B_traj_001157/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001156_B_traj_001157/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001293_A_traj_001293/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001293_A_traj_001293/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001293_A_traj_001293/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001293_B_traj_001294/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001293_B_traj_001294/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001293_B_traj_001294/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001428_A_traj_001428/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001428_A_traj_001428/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001428_A_traj_001428/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001428_B_traj_001429/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001428_B_traj_001429/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_001428_B_traj_001429/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002019_A_traj_002019/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002019_A_traj_002019/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002019_A_traj_002019/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002019_B_traj_002020/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002019_B_traj_002020/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002019_B_traj_002020/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002170_A_traj_002170/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002170_A_traj_002170/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002170_A_traj_002170/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002170_B_traj_002171/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002170_B_traj_002171/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260411_181558__pair_002170_B_traj_002171/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000294_A_traj_000294/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000294_A_traj_000294/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000294_A_traj_000294/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000294_B_traj_000295/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000294_B_traj_000295/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000294_B_traj_000295/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000698_A_traj_000698/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000698_A_traj_000698/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000698_A_traj_000698/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000698_B_traj_000699/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000698_B_traj_000699/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_000698_B_traj_000699/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_001596_A_traj_001596/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_001596_A_traj_001596/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_001596_A_traj_001596/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_001596_B_traj_001597/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_001596_B_traj_001597/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_001596_B_traj_001597/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_002260_A_traj_002260/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_002260_A_traj_002260/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_002260_A_traj_002260/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_002260_B_traj_002261/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_002260_B_traj_002261/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_002260_B_traj_002261/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_003301_A_traj_003301/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_003301_A_traj_003301/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_003301_A_traj_003301/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_003301_B_traj_003302/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_003301_B_traj_003302/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260413_221314__pair_003301_B_traj_003302/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000154_A_traj_000154/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000154_A_traj_000154/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000154_A_traj_000154/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000154_B_traj_000155/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000154_B_traj_000155/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000154_B_traj_000155/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000172_A_traj_000172/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000172_A_traj_000172/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000172_A_traj_000172/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000172_B_traj_000173/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000172_B_traj_000173/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_132732__pair_000172_B_traj_000173/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_205421__pair_000934_A_traj_000934/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_205421__pair_000934_A_traj_000934/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_205421__pair_000934_A_traj_000934/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_205421__pair_000934_B_traj_000935/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_205421__pair_000934_B_traj_000935/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260414_205421__pair_000934_B_traj_000935/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260415_212208__pair_000621_A_traj_000621/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260415_212208__pair_000621_A_traj_000621/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260415_212208__pair_000621_A_traj_000621/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260415_212208__pair_000621_B_traj_000622/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260415_212208__pair_000621_B_traj_000622/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260415_212208__pair_000621_B_traj_000622/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_001798_A_traj_001798/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_001798_A_traj_001798/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_001798_A_traj_001798/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_001798_B_traj_001799/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_001798_B_traj_001799/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_001798_B_traj_001799/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002491_A_traj_002491/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002491_A_traj_002491/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002491_A_traj_002491/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002491_B_traj_002492/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002491_B_traj_002492/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002491_B_traj_002492/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002778_A_traj_002778/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002778_A_traj_002778/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002778_A_traj_002778/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002778_B_traj_002779/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002778_B_traj_002779/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002778_B_traj_002779/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002855_A_traj_002855/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002855_A_traj_002855/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002855_A_traj_002855/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002855_B_traj_002856/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002855_B_traj_002856/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_002855_B_traj_002856/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_003390_A_traj_003390/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_003390_A_traj_003390/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_003390_A_traj_003390/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_003390_B_traj_003391/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_003390_B_traj_003391/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_003390_B_traj_003391/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004217_A_traj_004217/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004217_A_traj_004217/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004217_A_traj_004217/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004217_B_traj_004218/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004217_B_traj_004218/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004217_B_traj_004218/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004944_A_traj_004944/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004944_A_traj_004944/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004944_A_traj_004944/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004944_B_traj_004945/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004944_B_traj_004945/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_004944_B_traj_004945/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005292_A_traj_005292/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005292_A_traj_005292/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005292_A_traj_005292/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005292_B_traj_005293/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005292_B_traj_005293/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005292_B_traj_005293/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005324_A_traj_005324/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005324_A_traj_005324/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005324_A_traj_005324/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005324_B_traj_005325/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005324_B_traj_005325/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/equivalence_easy__run_20260422_172644__pair_005324_B_traj_005325/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260407_174321__traj_002355/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260407_174321__traj_002355/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260407_174321__traj_002355/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260407_174321__traj_002454/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260407_174321__traj_002454/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260407_174321__traj_002454/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_001504/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_001504/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_001504/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_002567/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_002567/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_002567/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_003021/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_003021/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_003021/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_004105/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_004105/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_004105/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_005386/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_005386/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260408_180248__traj_005386/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260409_134704__traj_000431/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260409_134704__traj_000431/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260409_134704__traj_000431/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_001167/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_001167/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_001167/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_001498/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_001498/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_001498/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_003759/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_003759/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_003759/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_003942/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_003942/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_003942/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_004594/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_004594/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_184831__traj_004594/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_204603__traj_000013/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_204603__traj_000013/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260410_204603__traj_000013/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260411_181558__traj_001520/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260411_181558__traj_001520/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260411_181558__traj_001520/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260411_181558__traj_003260/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260411_181558__traj_003260/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260411_181558__traj_003260/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000766/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000766/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000766/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000826/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000826/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000826/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000839/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000839/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_000839/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_002041/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_002041/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260413_221314__traj_002041/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_132732__traj_000031/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_132732__traj_000031/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_132732__traj_000031/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_205421__traj_000479/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_205421__traj_000479/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_205421__traj_000479/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_205421__traj_001251/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_205421__traj_001251/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260414_205421__traj_001251/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260422_172644__traj_000024/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260422_172644__traj_000024/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260422_172644__traj_000024/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260422_172644__traj_006125/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260422_172644__traj_006125/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_easy__run_20260422_172644__traj_006125/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_hard__run_20260427_001012__traj_000344/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_hard__run_20260427_001012__traj_000344/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/inverse_hard__run_20260427_001012__traj_000344/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260407_132406__traj_000496/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260407_132406__traj_000496/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260407_132406__traj_000496/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260410_184831__traj_001649/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260410_184831__traj_001649/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260410_184831__traj_001649/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler
diff --git a/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260422_172644__traj_003591/stage1_config.yaml b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260422_172644__traj_003591/stage1_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..529f90632f5b90bb1cfc9cfbe7bbaefebcb64f77
--- /dev/null
+++ b/inference/2026-09-23/outputs/sana_r50_local_v1/sc/raw/loop_easy__run_20260422_172644__traj_003591/stage1_config.yaml
@@ -0,0 +1,75 @@
+model:
+  model: SanaMSVideoCamCtrlStreaming_1600M_P1_D20
+  image_size: 352
+  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
+  mixed_precision: bf16
+  fp32_attention: true
+  multi_scale: true
+  attn_type: BidirectionalGDNTriton
+  softmax_every_n: 4
+  linear_head_dim: 112
+  conv_kernel_size: 4
+  k_conv_only: true
+  ffn_type: CachedGLUMBConvTemp
+  t_kernel_size: 3
+  mlp_acts:
+  - silu
+  - silu
+  - null
+  mlp_ratio: 3
+  use_pe: true
+  pos_embed_type: casual_wan_rope
+  qk_norm: true
+  cross_norm: true
+  class_dropout_prob: 0.0
+  chunk_size: 3
+  chunk_split_strategy: first_chunk_plus_one
+  cam_attn_compress: 1
+  init_cam_from_base: true
+  use_chunk_plucker_post_attn: true
+  chunk_plucker_channels: 48
+  chunk_plucker_post_attn_blocks: 20
+vae:
+  vae_type: LTX2VAE_diffusers_causal
+  vae_pretrained: /dev/shm/yifan_r50_sana_weights/SANA-WM_streaming/ltx2_causal_vae
+  weight_dtype: bfloat16
+  vae_latent_dim: 128
+  vae_downsample_rate: 32
+  vae_stride:
+  - 8
+  - 32
+  - 32
+  use_framewise_encoding: true
+  use_framewise_decoding: true
+  tile_sample_stride_num_frames: 64
+  tile_sample_min_num_frames: 96
+text_encoder:
+  text_encoder_name: gemma-2-2b-it
+  y_norm: true
+  y_norm_scale_factor: 0.01
+  model_max_length: 300
+  chi_prompt:
+  - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual
+    descriptions suitable for image generation. Evaluate the level of detail in the
+    user prompt:'
+  - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes,
+    textures, and spatial relationships to create vivid and concrete scenes.'
+  - '- If the prompt is already detailed, refine and enhance the existing details
+    slightly without overcomplicating.'
+  - 'Here are examples of how to transform or refine prompts:'
+  - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up
+    in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by
+    pots of blooming red flowers.'
+  - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at
+    dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing,
+    and a double-decker bus passing by towering glass skyscrapers.'
+  - 'Please generate only the enhanced description for the prompt below and avoid
+    including any additional commentary or evaluations:'
+  - 'User Prompt: '
+scheduler:
+  predict_flow_v: true
+  noise_schedule: linear_flow
+  pred_sigma: false
+  flow_shift: 9.95
+  inference_flow_shift: 8.0
+  vis_sampler: self_forcing_flow_euler