3D Task Summary - VDP / Perceptual Loss Screening

Short analysis

Best artifact-mining case so far: in-the-wild train views with clean held-out test views. It creates obvious smoky/exposure artifacts and is more useful than style or mild noise.
VDP heatmap weighted L1 has a small positive signal on wild-train reconstruction: +0.16 PSNR, +0.006 SSIM, -0.004 LPIPS. The visual gain is not yet a strong paper-level artifact fix.
VDP CSF/direct variants are mixed: sometimes improve JOD/LPIPS, but can hurt PSNR or spatial quality. Direct CVVDP should not be the main branch right now.
Teacher trajectory CVVDP finetune is consistent across LightGaussian, CompGS, and Distilled-3DGS on rendered path metrics, but held-out test-view gains are modest.
Aggressive LightGaussian compression on truck did not clearly amplify shimmer/popping; this is not a strong artifact source in the current setup.
Most reasonable next filter: keep heatmap-L1, search for stronger temporal/compression artifacts, and avoid scaling expensive CVVDP training until the visual artifact is obvious.

1. Artifact mining / task screening

Main purpose: find a 3D setting where plain per-view L1 is visibly insufficient.

Task	PSNR	SSIM	LPIPS	Read	Detail
g3_style_l1	24.609	0.8450	0.1757	Style-shifted train views; L1 baseline already reasonable.	open
g4_noisy_train_l1	24.356	0.8516	0.1777	Noisy train, clean test; degradation is visible but not severe.	open
g5_wild_train_l1	16.895	0.6899	0.3368	Wild exposure/WB/gamma/vignette train, clean test; strongest artifact source.	open

Wild-train L1 baseline: clearest artifacts

Aggressive LightGaussian compression

Variant	Size MB	PSNR	SSIM	LPIPS	Read
lg_vq_baseline_existing	41.3	24.559	0.8559	0.1836	No strong shimmer/popping amplification in this truck screening.
lg_vq_aggressive_r0p25_cb1024	55.7	24.613	0.8578	-	No strong shimmer/popping amplification in this truck screening.
lg_vq_aggressive_r0p10_cb512	62.0	24.623	0.8583	-	No strong shimmer/popping amplification in this truck screening.

2. VDP loss comparison on wild-train case

VDP losses still use the corrupted training references. This is a robustness screen, not clean-supervised restoration.

Method	PSNR	dPSNR	SSIM	dSSIM	LPIPS	LPIPS gain
L1 baseline	16.895	0.000	0.6899	0.0000	0.3368	0.0000
VDP heatmap weighted L1	17.058	+0.163	0.6958	+0.0060	0.3328	+0.0041
VDP CSF + L1	16.792	-0.103	0.6980	+0.0081	0.3258	+0.0111

Clean GT, L1, VDP heat-L1, VDP CSF+L1, amplified differences

Open detailed VDP report

3. Teacher trajectory CVVDP finetune

All three compact/distilled students were finetuned against a 3DGS teacher on rendered camera paths. Gains are consistent on teacher-path metrics; held-out test-view gains are modest.

Method	Test PSNR before	Test PSNR after	dPSNR	LPIPS before	LPIPS after	LPIPS gain	Path PSNR before	Path PSNR after	Path dPSNR	Detail
LightGaussian	25.511	25.558	+0.047	0.1518	0.1512	+0.0006	35.492	38.103	+2.611	open
CompGS	23.909	24.271	+0.362	0.2015	0.1950	+0.0065	27.833	29.449	+1.616	open
Distilled-3DGS	25.875	25.890	+0.015	0.1420	0.1423	-0.0003	30.854	31.949	+1.095	open

Open 6-way aggregate report

4. LightGaussian training variants

Earlier direct LG distillation comparison: heatmap-L1 is slightly positive; direct CVVDP hurts frame metrics and is not the current priority.

Variant report	Method	PSNR	SSIM	L1	JOD
heatmap	Baseline LG final VQ	24.818	0.8626	0.0342	8.838
heatmap	CVVDP heatmap LG distillation	24.923	0.8659	0.0335	8.861
heatmap	CVVDP heatmap LG final VQ	24.863	0.8633	0.0339	8.848
direct	Baseline LG final VQ	24.818	0.8626	0.0342	8.838
direct	CVVDP direct LG distillation	24.619	0.8553	0.0351	8.837
direct	CVVDP direct LG final VQ	24.561	0.8528	0.0355	8.824

Full video reports are summarized here but omitted from the lightweight public package.

5. Temporal diagnostic on truck trajectory

Baseline truck does not expose a very obvious temporal artifact. CVVDP dump channels and transient HF proxy are included for inspection.

Method	PSNR	SSIM	L1	Temporal L1 delta	CVVDP JOD
LG baseline final VQ	28.138	0.9411	0.0214	0.0241	8.262
LG ours heatmap L1 final VQ	28.203	0.9415	0.0211	0.0240	8.283
LG ours direct CVVDP final VQ	27.563	0.9316	0.0231	0.0262	8.327

Detail report links

Folder	Link
3d_task_screening_20260712	open
3d_task_vdp_compare_wild_20260712	open
trajectory_cvvdp_6way_truck	open
lightgaussian_traj_cvvdp_truck	open
compgs_traj_cvvdp_truck	open
distilled3dgs_traj_cvvdp_truck	open
compact3dgs_baselines_stage1	open
distilled_cvvdp_weight_stats	open