V5 — 4x Tensor Super-Resolution

Date: 2026-07-13. Source video: /local/scratch-2/fy277/temp/vdp_screening/input/cycle_view.mp4. Clip: 3.0s at 8 fps, resized to 384×256. Optimization iterations: L1/L2/CVVDP-heatmap L1 = 260; raw CVVDP = 70. Task: optimize the 4x downsampled RGB tensor directly, then bicubic-upsample to reference resolution. Losses: L1, L2, detached CVVDP raw-heatmap weighted L1, and direct raw CVVDP scalar.

1 · Full Metric Table

Name	Loss	JOD	ΔJOD	PSNR	ΔPSNR	SSIM	L1	L2	LPIPS	ΔLPIPS	Seconds
`raw_cvvdp`	-CVVDP JOD + 0.01 L1	7.885	+1.560	22.220	-0.215	0.595	0.05143	0.005998	0.6096	+0.0471	15.5
`l2`	L2	7.588	+1.263	24.186	+1.751	0.674	0.03871	0.003814	0.5172	-0.0453	0.3
`vdp_heat_l1`	CVVDP raw heatmap × L1	7.392	+1.067	23.797	+1.363	0.663	0.03780	0.004171	0.5235	-0.0389	0.3
`l1`	L1	7.331	+1.006	23.765	+1.330	0.670	0.03691	0.004202	0.5284	-0.0341	0.4
`bicubic`	none	6.325	+0.000	22.435	+0.000	0.627	0.04538	0.005709	0.5625	+0.0000	0.0

2 · Videos

Reference clip

4x downsampled input

Bicubic baseline

Optimized tensor · L1

Optimized tensor · L2

Optimized tensor · CVVDP heatmap L1

Optimized tensor · raw CVVDP scalar

Bicubic CVVDP raw heatmap

3 · Diagnostics

Frame comparison grid

Optimization curves

PSNR per frame

4 · Verdict

This run optimizes the 4x low-resolution tensor directly on a different 3-second video clip. Compared with bicubic, L1 changes JOD by +1.006, L2 by +1.263, CVVDP-heatmap L1 by +1.067, and raw CVVDP by +1.560. The raw CVVDP column is a direct differentiable scalar-loss probe; the heatmap-weighted variant keeps CVVDP detached and uses it only as spatial perceptual weighting.

5 · Config

{
  "source_video": "/local/scratch-2/fy277/temp/vdp_screening/input/cycle_view.mp4",
  "work_dir": "/local/scratch-2/fy277/temp/vdp_screening/work/v5_cycle_view_superres_tensor_opt_20260713",
  "seconds": 3.0,
  "start": 0.0,
  "fps": 8,
  "frames": 24,
  "width": 384,
  "height": 256,
  "downsample_factor": 4,
  "low_width": 96,
  "low_height": 64,
  "steps": 260,
  "raw_steps": 70,
  "lr": 0.08,
  "raw_lr": 0.025,
  "device": "cuda",
  "cuda_visible_devices": "0"
}