V4 - Frame-wise Flicker Removal
Date: 2026-07-13. Source video: /local/scratch-2/fy277/temp/vdp_screening/input/cycle_view.mp4. Clip: 3.0s at 8 fps, 384x256. Optimization iterations: 200 for each correction run. Correction model: per-frame per-channel affine tensor.
0 - Label Legend
CVVDP usage: the whole video clip is passed to CVVDP to compute JOD and, for cvvdp_hm losses, a detached raw CVVDP heatmap bank. Optimization does not backpropagate through the CVVDP scalar here; it optimizes L1 or heatmap-weighted L1.
alpha_only: learn per-frame, per-channel multiplicative gain only: out_t = alpha_t * input_t.
alpha_beta: learn per-frame, per-channel gain and bias: out_t = alpha_t * input_t + beta_t.
cvvdp_hm_a1: detached CVVDP heatmap weighted L1 with heatmap weight strength 1.
cvvdp_hm_a4: detached CVVDP heatmap weighted L1 with heatmap weight strength 4.
1 - Flicker Inputs
| Input | JOD | PSNR | SSIM | L1 | LPIPS |
|---|
cycle_view_global_brightness_amp0p20 | 6.334 | 42.114 | 0.982 | 0.04568 | 0.0072 |
cycle_view_color_flicker_amp0p50 | 2.340 | 21.524 | 0.918 | 0.08399 | 0.1618 |
2 - Results
cycle_view_global_brightness_amp0p20
| Model | Loss | JOD | PSNR | SSIM | L1 | LPIPS |
|---|
alpha_only | cvvdp_hm_a1 | 9.774 | 64.026 | 0.999 | 0.00088 | 0.0002 |
alpha_only | l1 | 9.667 | 63.630 | 0.999 | 0.00097 | 0.0003 |
alpha_only | cvvdp_hm_a4 | 9.701 | 63.232 | 0.999 | 0.00114 | 0.0003 |
alpha_beta | cvvdp_hm_a4 | 9.168 | 58.213 | 0.999 | 0.00402 | 0.0013 |
alpha_beta | l1 | 9.315 | 58.129 | 0.999 | 0.00407 | 0.0011 |
cycle_view_global_brightness_amp0p20 input
cycle_view_color_flicker_amp0p50
| Model | Loss | JOD | PSNR | SSIM | L1 | LPIPS |
|---|
alpha_only | cvvdp_hm_a1 | 8.892 | 41.051 | 0.991 | 0.00340 | 0.0112 |
alpha_only | cvvdp_hm_a4 | 8.881 | 40.988 | 0.991 | 0.00336 | 0.0113 |
alpha_only | l1 | 8.877 | 40.860 | 0.991 | 0.00334 | 0.0113 |
alpha_beta | cvvdp_hm_a4 | 8.311 | 39.280 | 0.991 | 0.00643 | 0.0118 |
alpha_beta | l1 | 8.708 | 38.846 | 0.991 | 0.00612 | 0.0120 |
cycle_view_color_flicker_amp0p50 input
3 - Verdict
Cycle-view flicker-removal rerun. The synthetic inputs are generated from the new reference clip, then a per-frame affine correction is fit with L1 or detached CVVDP-heatmap-weighted L1. All embedded videos are H.264/yuv420p/faststart.
4 - Config
{
"source_video": "/local/scratch-2/fy277/temp/vdp_screening/input/cycle_view.mp4",
"work_dir": "/local/scratch-2/fy277/temp/vdp_screening/work/v1_v4_cycle_view_20260713",
"v1_report": "/local/scratch-2/fy277/temp/vdp_screening/reports/v1_compression_2026-07-12_cycle_view_new",
"v4_report": "/local/scratch-2/fy277/temp/vdp_screening/reports/v4_flicker_removal_2026-07-12_cycle_view_new",
"seconds": 3.0,
"start": 0.0,
"fps": 8,
"width": 384,
"v4_iters": 200,
"only": "all",
"height": 256,
"frames": 24
}