Toys
Amplitude L2 · LR 0.2
final JOD 9.714 · PSNR 39.13 dB · SSIM 0.9103
sRGB Charbonnier · LR 0.05
final JOD 9.389 · PSNR 38.88 dB · SSIM 0.9075
2D HDR: learning rate was the missing variable
Four representative scenes (two darker, two brighter), native ROI, no padding, 8 hard-binary frames, 300 iterations. Each loss gets one global LR selected by four-scene mean JOD.
0.2L2 selected LR
9.581L2 mean JOD
0.05Charb selected LR
9.314Charb mean JOD
LR sweep
| Loss | LR | mean JOD | mean PSNR | mean SSIM | |
|---|---|---|---|---|---|
| Amplitude L2 | 0.005 | 8.906 | 25.84 | 0.6239 | |
| Amplitude L2 | 0.05 | 9.449 | 30.91 | 0.8063 | |
| Amplitude L2 | 0.2 | 9.581 | 32.78 | 0.8496 | selected |
| sRGB Charbonnier | 0.005 | 9.022 | 27.98 | 0.7101 | |
| sRGB Charbonnier | 0.05 | 9.314 | 32.32 | 0.8364 | selected |
| sRGB Charbonnier | 0.1 | 9.305 | 32.62 | 0.8397 |
Selected-LR scene results
| Scene | Amplitude L2 JOD | sRGB Charbonnier JOD |
|---|---|---|
| Toys | 9.714 | 9.389 |
| Rushmore | 9.460 | 8.973 |
| Water | 9.651 | 9.520 |
| Castle | 9.497 | 9.374 |
The iteration snapshots below show whether the gain comes from continuing convergence rather than a different forward model.
Rushmore
Amplitude L2 · LR 0.2
final JOD 9.460 · PSNR 30.89 dB · SSIM 0.7938
sRGB Charbonnier · LR 0.05
final JOD 8.973 · PSNR 29.96 dB · SSIM 0.7684
Water
Amplitude L2 · LR 0.2
final JOD 9.651 · PSNR 30.68 dB · SSIM 0.8658
sRGB Charbonnier · LR 0.05
final JOD 9.520 · PSNR 30.21 dB · SSIM 0.8529
Castle
Amplitude L2 · LR 0.2
final JOD 9.497 · PSNR 30.41 dB · SSIM 0.8283
sRGB Charbonnier · LR 0.05
final JOD 9.374 · PSNR 30.24 dB · SSIM 0.8170
Interpretation
This validates the postdoc-facing control: the earlier weak L2/Charbonnier result was primarily an optimizer-budget mismatch. Large LR reaches a qualitatively cleaner hard reconstruction by 100–300 iterations on both dark and bright HDR scenes. Any new perceptual loss must therefore beat the independently tuned baselines, not LR 5e-3.