Controlled HDR ablation · one seed

Where does the gain come from: loss formula or display domain?

Four matched runs isolate two choices: amplitude versus sRGB residuals, and L2 versus Charbonnier. Toys represents a dark HDR distribution; Castle represents a bright one.

Native ROINo paddingLR 5e−3300 iterations8-frame GumbelHard evaluationSame initialization

Primary comparison

Domain-only effect

Each card is sRGB-domain loss minus the same formula in amplitude. Positive JOD/PSNR/SSIM is better.

dark HDR · domain only

Toys: sRGB L2 − amplitude L2

+0.149JOD +0.91 dBPSNR +0.0219SSIM

bright HDR · domain only

Castle: sRGB L2 − amplitude L2

+0.068JOD +0.16 dBPSNR -0.0026SSIM

dark HDR · domain only

Toys: sRGB Charb − amplitude Charb

-0.021JOD -0.17 dBPSNR -0.0100SSIM

bright HDR · domain only

Castle: sRGB Charb − amplitude Charb

+0.029JOD +0.06 dBPSNR -0.0001SSIM

Formula-only comparison

Robust residual effect

Amplitude Charbonnier minus amplitude L2 keeps the domain fixed and changes only the penalty shape.

dark HDR · formula only

Toys: amplitude Charb − amplitude L2

+0.120JOD +3.86 dBPSNR +0.1154SSIM

bright HDR · formula only

Castle: amplitude Charb − amplitude L2

+0.248JOD +1.96 dBPSNR +0.0970SSIM
Result: sRGB conversion helps L2 substantially on dark Toys, but barely helps bright Castle. It does not improve Charbonnier in Toys and is nearly neutral in Castle. In these two matched runs, the larger and more consistent gain comes from the Charbonnier penalty itself—not from converting residuals to sRGB.

dark HDR · median 1.9 nits

Toys

Toys target

Ground truth

Per-image q99.9-normalized HDR target

Amplitude L2 reconstruction for Toys

Amplitude L2

JOD 9.083 · PSNR 28.08 dB · SSIM 0.6451 · L2 0.001555

sRGB L2 reconstruction for Toys

sRGB L2

JOD 9.232 · PSNR 28.99 dB · SSIM 0.6670 · L2 0.001262

Amplitude Charbonnier reconstruction for Toys

Amplitude Charbonnier

JOD 9.203 · PSNR 31.94 dB · SSIM 0.7605 · L2 0.000640

sRGB Charbonnier reconstruction for Toys

sRGB Charbonnier

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

Loss at 300JOD ↑PSNR ↑SSIM ↑sRGB L2 ↓
Amplitude L29.08328.080.64510.001555
sRGB L29.23228.990.66700.001262
Amplitude Charbonnier9.20331.940.76050.000640
sRGB Charbonnier9.18231.770.75050.000665
Metric curves for Toys
Hard reconstruction metrics, sampled every 20 iterations.

bright HDR · median 184.6 nits

Castle

Castle target

Ground truth

Per-image q99.9-normalized HDR target

Amplitude L2 reconstruction for Castle

Amplitude L2

JOD 8.772 · PSNR 24.71 dB · SSIM 0.6018 · L2 0.003378

sRGB L2 reconstruction for Castle

sRGB L2

JOD 8.840 · PSNR 24.87 dB · SSIM 0.5992 · L2 0.003256

Amplitude Charbonnier reconstruction for Castle

Amplitude Charbonnier

JOD 9.020 · PSNR 26.67 dB · SSIM 0.6988 · L2 0.002154

sRGB Charbonnier reconstruction for Castle

sRGB Charbonnier

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

Loss at 300JOD ↑PSNR ↑SSIM ↑sRGB L2 ↓
Amplitude L28.77224.710.60180.003378
sRGB L28.84024.870.59920.003256
Amplitude Charbonnier9.02026.670.69880.002154
sRGB Charbonnier9.04926.730.69870.002124
Metric curves for Castle
Hard reconstruction metrics, sampled every 20 iterations.