H2-conv (3.17) step 5000: eval tables

image_video_conv/step5000/README.md

@@ -9,3 +9,17 @@ after the splat, splat_direct, 8-head mid attention, conv_low_in). Trained on im
 - `s5000_*.md`: eval tables (200 ImageNet val images; 20 OpenVid held-out clips, stride 8 and 2).
 - `launcher.sh`: the training flags (paths removed).
 
+
+Eval (200 ImageNet val images; 20 OpenVid clips):
+- s5000_img:
+    | 200 | 28.65 | 0.833 | 0.100 | 22.71 | 16420 |
+- s5000ema_img:
+    | 200 | 29.25 | 0.847 | 0.097 | 20.72 | 16420 |
+- s5000_st8:
+    | anchor | 30.77 | 0.895 | 0.080 | - |
+    | target, full state | 26.41 | 0.799 | 0.164 | 28626 |
+    | target, L = 0 | 25.68 | 0.785 | 0.175 | 24882 |
+- s5000_st2:
+    | anchor | 30.90 | 0.896 | 0.080 | - |
+    | target, full state | 28.38 | 0.855 | 0.120 | 28626 |
+    | target, L = 0 | 27.80 | 0.843 | 0.125 | 24882 |

image_video_conv/step5000/tables/s5000_img.md

@@ -0,0 +1,9 @@
+# Grain-Prior C-D v1 eval, step 5000
+
+query_init=cond geom_bias=True memory=True image ledger=16420 full state=28626
+
+## Images
+
+| n | PSNR | SSIM | LPIPS | rFID | numbers |
+|---|---|---|---|---|---|
+| 200 | 28.65 | 0.833 | 0.100 | 22.71 | 16420 |

image_video_conv/step5000/tables/s5000_st2.md

@@ -0,0 +1,21 @@
+# Grain-Prior C-D v1 eval, step 5000
+
+query_init=cond geom_bias=True memory=True image ledger=16420 full state=28626
+
+## video manifest (20 clips, payload 28626)
+
+| decode | PSNR | SSIM | LPIPS | numbers |
+|---|---|---|---|---|
+| anchor | 30.90 | 0.896 | 0.080 | - |
+| target, full state | 28.38 | 0.855 | 0.120 | 28626 |
+| target, L = 0 | 27.80 | 0.843 | 0.125 | 24882 |
+| target, L = 52 | 28.38 | 0.855 | 0.120 | 28626 |
+| novel view: grains + camera only (z = c_0, L = 0) | 20.26 | 0.591 | 0.249 | 24882 |
+| copy anchor (no model) | 24.81 | 0.752 | 0.115 | - |
+| copy last frame (no model) | 27.56 | 0.831 | 0.070 | - |
+
+target by record (full state), PSNR: 29.17 / 28.34 / 27.63
+
+gates: L0 - copy-last = 0.24 dB; novel view - copy anchor = -4.55 dB
+
+init_cos=0.705 base_cos_pass0=0.810 base_cos_pass1=0.819 base_cos_final=0.838 feat_cos=0.842 support_frac=1.000 psnr_base=27.804 xi_rot_deg=0.606 xi_trans=0.362 fov_deg=60.770 rho_std=0.021 mem_delta=0.031 sigma_mean=0.755 psnr_copy_last=27.560 splat_dropped_frac=0.000

image_video_conv/step5000/tables/s5000_st8.md

@@ -0,0 +1,21 @@
+# Grain-Prior C-D v1 eval, step 5000
+
+query_init=cond geom_bias=True memory=True image ledger=16420 full state=28626
+
+## video manifest (20 clips, payload 28626)
+
+| decode | PSNR | SSIM | LPIPS | numbers |
+|---|---|---|---|---|
+| anchor | 30.77 | 0.895 | 0.080 | - |
+| target, full state | 26.41 | 0.799 | 0.164 | 28626 |
+| target, L = 0 | 25.68 | 0.785 | 0.175 | 24882 |
+| target, L = 52 | 26.41 | 0.799 | 0.164 | 28626 |
+| novel view: grains + camera only (z = c_0, L = 0) | 17.51 | 0.439 | 0.350 | 24882 |
+| copy anchor (no model) | 19.51 | 0.573 | 0.221 | - |
+| copy last frame (no model) | 21.73 | 0.658 | 0.157 | - |
+
+target by record (full state), PSNR: 27.26 / 26.37 / 25.61
+
+gates: L0 - copy-last = 3.95 dB; novel view - copy anchor = -2.00 dB
+
+init_cos=0.701 base_cos_pass0=0.795 base_cos_pass1=0.807 base_cos_final=0.830 feat_cos=0.835 support_frac=1.000 psnr_base=25.698 xi_rot_deg=0.842 xi_trans=0.479 fov_deg=60.703 rho_std=0.022 mem_delta=0.036 sigma_mean=0.754 psnr_copy_last=21.728 splat_dropped_frac=0.002

image_video_conv/step5000/tables/s5000ema_img.md

@@ -0,0 +1,9 @@
+# Grain-Prior C-D v1 eval, step 5000
+
+query_init=cond geom_bias=True memory=True image ledger=16420 full state=28626
+
+## Images
+
+| n | PSNR | SSIM | LPIPS | rFID | numbers |
+|---|---|---|---|---|---|
+| 200 | 29.25 | 0.847 | 0.097 | 20.72 | 16420 |