H2-conv (3.17) step 12000: eval tables

image_video_conv/step12000/README.md

@@ -9,3 +9,17 @@ after the splat, splat_direct, 8-head mid attention, conv_low_in). Trained on im
 - `s12000_*.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):
+- s12000_img:
+    | 200 | 29.63 | 0.855 | 0.092 | 20.88 | 16420 |
+- s12000ema_img:
+    | 200 | 30.01 | 0.864 | 0.086 | 21.23 | 16420 |
+- s12000_st8:
+    | anchor | 32.15 | 0.916 | 0.072 | - |
+    | target, full state | 27.09 | 0.823 | 0.144 | 28626 |
+    | target, L = 0 | 26.91 | 0.821 | 0.149 | 24882 |
+- s12000_st2:
+    | anchor | 32.28 | 0.917 | 0.072 | - |
+    | target, full state | 29.26 | 0.876 | 0.105 | 28626 |
+    | target, L = 0 | 29.17 | 0.873 | 0.107 | 24882 |

image_video_conv/step12000/tables/s12000_img.md

@@ -0,0 +1,9 @@
+# Grain-Prior C-D v1 eval, step 12000
+
+query_init=cond geom_bias=True memory=True image ledger=16420 full state=28626
+
+## Images
+
+| n | PSNR | SSIM | LPIPS | rFID | numbers |
+|---|---|---|---|---|---|
+| 200 | 29.63 | 0.855 | 0.092 | 20.88 | 16420 |

image_video_conv/step12000/tables/s12000_st2.md

@@ -0,0 +1,21 @@
+# Grain-Prior C-D v1 eval, step 12000
+
+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 | 32.28 | 0.917 | 0.072 | - |
+| target, full state | 29.26 | 0.876 | 0.105 | 28626 |
+| target, L = 0 | 29.17 | 0.873 | 0.107 | 24882 |
+| target, L = 52 | 29.26 | 0.876 | 0.105 | 28626 |
+| novel view: grains + camera only (z = c_0, L = 0) | 20.24 | 0.560 | 0.280 | 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: 30.10 / 29.25 / 28.44
+
+gates: L0 - copy-last = 1.61 dB; novel view - copy anchor = -4.57 dB
+
+init_cos=0.705 base_cos_pass0=0.810 base_cos_pass1=0.818 base_cos_final=0.839 feat_cos=0.843 support_frac=1.000 psnr_base=29.178 xi_rot_deg=0.691 xi_trans=0.372 fov_deg=62.890 rho_std=0.021 mem_delta=0.033 sigma_mean=0.788 psnr_copy_last=27.560 splat_dropped_frac=0.000

image_video_conv/step12000/tables/s12000_st8.md

@@ -0,0 +1,21 @@
+# Grain-Prior C-D v1 eval, step 12000
+
+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 | 32.15 | 0.916 | 0.072 | - |
+| target, full state | 27.09 | 0.823 | 0.144 | 28626 |
+| target, L = 0 | 26.91 | 0.821 | 0.149 | 24882 |
+| target, L = 52 | 27.09 | 0.823 | 0.144 | 28626 |
+| novel view: grains + camera only (z = c_0, L = 0) | 16.96 | 0.378 | 0.410 | 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: 28.07 / 26.98 / 26.21
+
+gates: L0 - copy-last = 5.18 dB; novel view - copy anchor = -2.54 dB
+
+init_cos=0.700 base_cos_pass0=0.798 base_cos_pass1=0.807 base_cos_final=0.831 feat_cos=0.835 support_frac=1.000 psnr_base=26.913 xi_rot_deg=0.950 xi_trans=0.489 fov_deg=62.812 rho_std=0.021 mem_delta=0.037 sigma_mean=0.787 psnr_copy_last=21.728 splat_dropped_frac=0.002

image_video_conv/step12000/tables/s12000ema_img.md

@@ -0,0 +1,9 @@
+# Grain-Prior C-D v1 eval, step 12000
+
+query_init=cond geom_bias=True memory=True image ledger=16420 full state=28626
+
+## Images
+
+| n | PSNR | SSIM | LPIPS | rFID | numbers |
+|---|---|---|---|---|---|
+| 200 | 30.01 | 0.864 | 0.086 | 21.23 | 16420 |