CVVDP 0.6.0 · controlled heatmap ablation

Six ways to redistribute perceptual error

Two dark and two bright HDR scenes compare six new CVVDP heatmap constructions against plain sRGB Charbonnier and the existing CVVDP luma-heatmap × Charbonnier loss.

2× RTX 4090Native ROINo paddingLR 5e−3300 iterationsSame seedHard evaluationα 0.25

Bright-scene result

Luminance stratification wins

Deltas below are two-scene bright averages. The existing heatmap hurts Water/Castle; stratifying the same CVVDP heatmap within luminance bands reverses the sign on every metric.

Current − Charbonnier

Existing luma heatmap

-0.022JOD -0.23 dBPSNR -0.0180SSIM

D2 − Charbonnier

Luminance-stratified

+0.040JOD +0.15 dBPSNR +0.0051SSIM

D2 − Current

Replacement gain

+0.062JOD +0.38 dBPSNR +0.0230SSIM

Main result: on bright Water and Castle, D2 improves the Charbonnier mean by +0.040 JOD, +0.15 dB and +0.0051 SSIM. The original heatmap instead loses 0.23 dB and 0.0180 SSIM. On dark scenes, the original heatmap remains strongest because of its very large Toys gain; one normalization is therefore not optimal for both brightness regimes.

Dark mean · Toys + Type

Method	JOD ↑	PSNR ↑	SSIM ↑	sRGB L2 ↓
Charbonnier	8.826	28.07	0.5937	0.002161
Current luma heatmap × Charb	8.954	29.50	0.6241	0.001940
D1 · Log-budget	8.904	28.95	0.6126	0.001964
D2 · Luminance-stratified	8.864	28.04	0.5917	0.002095
D3 · Local-relative	8.813	27.95	0.5913	0.002216
D4 · Residual-novelty	8.847	28.27	0.5966	0.002103
D5 · Metric-warp	8.877	28.52	0.6034	0.002051
D6 · Hybrid	8.889	28.74	0.6079	0.002007

Bright mean · Water + Castle

Method	JOD ↑	PSNR ↑	SSIM ↑	sRGB L2 ↓
Charbonnier	9.119	26.80	0.7209	0.002089
Current luma heatmap × Charb	9.097	26.57	0.7029	0.002203
D1 · Log-budget	9.122	26.81	0.7191	0.002087
D2 · Luminance-stratified	9.160	26.95	0.7260	0.002020
D3 · Local-relative	9.127	26.78	0.7212	0.002099
D4 · Residual-novelty	9.107	26.64	0.7107	0.002170
D5 · Metric-warp	9.123	26.82	0.7207	0.002079
D6 · Hybrid	9.122	26.82	0.7201	0.002081

Implemented formulas

Six controlled designs

All six use the same detached CVVDP raw heatmap and the same sRGB Charbonnier residual. Only heatmap normalization and insertion point change.

D1 · Log-budget

z = clip(std(log(H+ε)), −2, 2)w = N[exp(0.25 · M · z)]

Stable, mean-one weak redistribution.

D2 · Luminance-stratified

z = robust-standardize(log H | luminance bin)w = N[exp(0.25 · z)]

Compares visibility within similar luminance levels.

D3 · Local-relative

z = std(log H − Gaussianσ=16(log H))w = N[exp(0.25 · M · z)]

Emphasizes peaks relative to their neighborhood.

D4 · Residual-novelty

z = std(zH − βzR)β = Cov(zH,zR) / Var(zR)

Removes the heatmap component explained by residual size.

D5 · Metric-warp

w = log-budget weightL = mean ρ(√w · r)

Stretches local residual distance inside Charbonnier.

D6 · Hybrid

γ = 0.5L = mean w^γ ρ(w^((1−γ)/2) · r)

Mixes outside weighting and inside distance warp.

dark HDR · median 1.9 nits

Toys

Toys ground truth

Ground truth

q99.9-normalized HDR target

Charbonnier reconstruction for Toys

Charbonnier

JOD 9.182 · PSNR 31.77 dB · SSIM 0.7505 · L2 0.000665

Current luma heatmap × Charb reconstruction for Toys

Current luma heatmap × Charb

JOD 9.410 · PSNR 34.48 dB · SSIM 0.8164 · L2 0.000357

D1 · Log-budget reconstruction for Toys

D1 · Log-budget

JOD 9.289 · PSNR 33.29 dB · SSIM 0.7868 · L2 0.000468

D2 · Luminance-stratified reconstruction for Toys

D2 · Luminance-stratified

JOD 9.223 · PSNR 31.50 dB · SSIM 0.7454 · L2 0.000708

D3 · Local-relative reconstruction for Toys

D3 · Local-relative

JOD 9.201 · PSNR 31.65 dB · SSIM 0.7482 · L2 0.000684

D4 · Residual-novelty reconstruction for Toys

D4 · Residual-novelty

JOD 9.206 · PSNR 32.07 dB · SSIM 0.7604 · L2 0.000621

D5 · Metric-warp reconstruction for Toys

D5 · Metric-warp

JOD 9.249 · PSNR 32.55 dB · SSIM 0.7692 · L2 0.000555

D6 · Hybrid reconstruction for Toys

D6 · Hybrid

JOD 9.270 · PSNR 32.92 dB · SSIM 0.7779 · L2 0.000510

Toys · final hard metrics

Method	JOD ↑	PSNR ↑	SSIM ↑	sRGB L2 ↓
Charbonnier	9.182	31.77	0.7505	0.000665
Current luma heatmap × Charb	9.410	34.48	0.8164	0.000357
D1 · Log-budget	9.289	33.29	0.7868	0.000468
D2 · Luminance-stratified	9.223	31.50	0.7454	0.000708
D3 · Local-relative	9.201	31.65	0.7482	0.000684
D4 · Residual-novelty	9.206	32.07	0.7604	0.000621
D5 · Metric-warp	9.249	32.55	0.7692	0.000555
D6 · Hybrid	9.270	32.92	0.7779	0.000510

Metric curves for Toys — Hard reconstruction metrics recorded every 20 iterations.

dark HDR · median 2.7 nits

Type

Type ground truth

Ground truth

q99.9-normalized HDR target

Charbonnier reconstruction for Type

Charbonnier

JOD 8.471 · PSNR 24.37 dB · SSIM 0.4369 · L2 0.003658

Current luma heatmap × Charb reconstruction for Type

Current luma heatmap × Charb

JOD 8.499 · PSNR 24.53 dB · SSIM 0.4319 · L2 0.003523

D1 · Log-budget reconstruction for Type

D1 · Log-budget

JOD 8.519 · PSNR 24.61 dB · SSIM 0.4384 · L2 0.003460

D2 · Luminance-stratified reconstruction for Type

D2 · Luminance-stratified

JOD 8.506 · PSNR 24.58 dB · SSIM 0.4381 · L2 0.003483

D3 · Local-relative reconstruction for Type

D3 · Local-relative

JOD 8.425 · PSNR 24.26 dB · SSIM 0.4344 · L2 0.003748

D4 · Residual-novelty reconstruction for Type

D4 · Residual-novelty

JOD 8.488 · PSNR 24.46 dB · SSIM 0.4328 · L2 0.003585

D5 · Metric-warp reconstruction for Type

D5 · Metric-warp

JOD 8.504 · PSNR 24.50 dB · SSIM 0.4375 · L2 0.003546

D6 · Hybrid reconstruction for Type

D6 · Hybrid

JOD 8.509 · PSNR 24.56 dB · SSIM 0.4380 · L2 0.003503

Type · final hard metrics

Method	JOD ↑	PSNR ↑	SSIM ↑	sRGB L2 ↓
Charbonnier	8.471	24.37	0.4369	0.003658
Current luma heatmap × Charb	8.499	24.53	0.4319	0.003523
D1 · Log-budget	8.519	24.61	0.4384	0.003460
D2 · Luminance-stratified	8.506	24.58	0.4381	0.003483
D3 · Local-relative	8.425	24.26	0.4344	0.003748
D4 · Residual-novelty	8.488	24.46	0.4328	0.003585
D5 · Metric-warp	8.504	24.50	0.4375	0.003546
D6 · Hybrid	8.509	24.56	0.4380	0.003503

Metric curves for Type — Hard reconstruction metrics recorded every 20 iterations.

bright HDR · median 76.2 nits

Water

Water ground truth

Ground truth

q99.9-normalized HDR target

Charbonnier reconstruction for Water

Charbonnier

JOD 9.190 · PSNR 26.87 dB · SSIM 0.7431 · L2 0.002054

Current luma heatmap × Charb reconstruction for Water

Current luma heatmap × Charb

JOD 9.198 · PSNR 26.79 dB · SSIM 0.7359 · L2 0.002093

D1 · Log-budget reconstruction for Water

D1 · Log-budget

JOD 9.201 · PSNR 26.93 dB · SSIM 0.7435 · L2 0.002028

D2 · Luminance-stratified reconstruction for Water

D2 · Luminance-stratified

JOD 9.218 · PSNR 27.05 dB · SSIM 0.7487 · L2 0.001972

D3 · Local-relative reconstruction for Water

D3 · Local-relative

JOD 9.202 · PSNR 26.87 dB · SSIM 0.7440 · L2 0.002055

D4 · Residual-novelty reconstruction for Water

D4 · Residual-novelty

JOD 9.189 · PSNR 26.76 dB · SSIM 0.7366 · L2 0.002109

D5 · Metric-warp reconstruction for Water

D5 · Metric-warp

JOD 9.200 · PSNR 26.92 dB · SSIM 0.7438 · L2 0.002031

D6 · Hybrid reconstruction for Water

D6 · Hybrid

JOD 9.197 · PSNR 26.93 dB · SSIM 0.7438 · L2 0.002028

Water · final hard metrics

Method	JOD ↑	PSNR ↑	SSIM ↑	sRGB L2 ↓
Charbonnier	9.190	26.87	0.7431	0.002054
Current luma heatmap × Charb	9.198	26.79	0.7359	0.002093
D1 · Log-budget	9.201	26.93	0.7435	0.002028
D2 · Luminance-stratified	9.218	27.05	0.7487	0.001972
D3 · Local-relative	9.202	26.87	0.7440	0.002055
D4 · Residual-novelty	9.189	26.76	0.7366	0.002109
D5 · Metric-warp	9.200	26.92	0.7438	0.002031
D6 · Hybrid	9.197	26.93	0.7438	0.002028

Metric curves for Water — Hard reconstruction metrics recorded every 20 iterations.

bright HDR · median 184.6 nits

Castle

Castle ground truth

Ground truth

q99.9-normalized HDR target

Charbonnier reconstruction for Castle

Charbonnier

JOD 9.049 · PSNR 26.73 dB · SSIM 0.6987 · L2 0.002124

Current luma heatmap × Charb reconstruction for Castle

Current luma heatmap × Charb

JOD 8.996 · PSNR 26.36 dB · SSIM 0.6700 · L2 0.002313

D1 · Log-budget reconstruction for Castle

D1 · Log-budget

JOD 9.043 · PSNR 26.69 dB · SSIM 0.6948 · L2 0.002145

D2 · Luminance-stratified reconstruction for Castle

D2 · Luminance-stratified

JOD 9.101 · PSNR 26.85 dB · SSIM 0.7032 · L2 0.002067

D3 · Local-relative reconstruction for Castle

D3 · Local-relative

JOD 9.051 · PSNR 26.69 dB · SSIM 0.6985 · L2 0.002143

D4 · Residual-novelty reconstruction for Castle

D4 · Residual-novelty

JOD 9.024 · PSNR 26.51 dB · SSIM 0.6849 · L2 0.002232

D5 · Metric-warp reconstruction for Castle

D5 · Metric-warp

JOD 9.047 · PSNR 26.72 dB · SSIM 0.6976 · L2 0.002127

D6 · Hybrid reconstruction for Castle

D6 · Hybrid

JOD 9.048 · PSNR 26.71 dB · SSIM 0.6963 · L2 0.002134

Castle · final hard metrics

Method	JOD ↑	PSNR ↑	SSIM ↑	sRGB L2 ↓
Charbonnier	9.049	26.73	0.6987	0.002124
Current luma heatmap × Charb	8.996	26.36	0.6700	0.002313
D1 · Log-budget	9.043	26.69	0.6948	0.002145
D2 · Luminance-stratified	9.101	26.85	0.7032	0.002067
D3 · Local-relative	9.051	26.69	0.6985	0.002143
D4 · Residual-novelty	9.024	26.51	0.6849	0.002232
D5 · Metric-warp	9.047	26.72	0.6976	0.002127
D6 · Hybrid	9.048	26.71	0.6963	0.002134

Metric curves for Castle — Hard reconstruction metrics recorded every 20 iterations.