8-frame Gumbel binary CGH · 300 iterations · hard reconstruction

PU21-MS-SSIM

A controlled two-step ablation: first test MS-SSIM in normalized linear intensity, then add fixed, display-referred PU21 before the same MS-SSIM objective. Each loss uses its own globally selected learning rate.

Outcome. Linear MS-SSIM versus amplitude L2: -1.025 mean JOD. PU21 contribution over linear MS-SSIM: +0.393 mean JOD. PU21-MS-SSIM versus the requested 9.372 target: -1.132 JOD.

12-scene mean JOD

Independent LR selection

Aggregate metrics

LossJODPSNRSSIM
Amplitude L28.87225.910.6283
Charbonnier8.95327.860.7004
Direct CVVDP9.06518.080.3759
Luma heatmap × Charb9.08128.420.7088
Linear MS-SSIM7.84723.440.6873
PU21-MS-SSIM8.24019.840.7261

Physical mapping

PU21 receives reconstruction and target intensity after the exact active CVVDP display photometry. No scene-wise normalization is applied inside either loss. PU normalization is fixed at 10,000 nit.

{
  "display_name": "standard_4k",
  "eotf": "sRGB",
  "y_peak_nits": 200.0,
  "contrast": 1000.0,
  "e_ambient_lux": 250.0,
  "k_reflection": 0.005,
  "y_black_nits": 0.2,
  "y_reflection_nits": 0.3978873577297384,
  "mapping": "L_nits=(200-0.2)*I_linear+0.2+0.39788735773"
}

Maximum observed clipping fraction: lower 0.000e+00; upper 0.000e+00.

Per-scene final hard JOD

SceneAmplitude L2CharbonnierDirect CVVDPLuma heatmap × CharbLinear MS-SSIMPU21-MS-SSIMPU−Linear
Anime28.4488.1009.0418.7747.8038.168+0.365
Bloom8.8739.0808.9449.1137.6997.934+0.235
Castle8.7729.0498.8018.9967.5358.070+0.535
Glade8.7458.8419.0168.8457.5888.299+0.711
Lounge9.1069.3329.3329.4718.5048.719+0.215
Meridian9.0589.2279.1599.3027.8398.204+0.365
Rushmore8.8198.6678.8938.6917.4578.027+0.570
Showgirl8.8238.8249.1339.1718.0348.303+0.269
Snow9.2969.4759.2599.5028.8218.783-0.038
Toys9.0839.1829.3799.4107.7078.268+0.561
Type8.4868.4718.8248.4997.3377.781+0.444
Water8.9529.1908.9989.1987.8378.322+0.485

Native ROI · hard binary

Anime2

Target

Linear MS-SSIM

JOD 7.803 · PSNR 21.26 · SSIM 0.6764

PU21-MS-SSIM

JOD 8.168 · PSNR 17.71 · SSIM 0.7110

Native ROI · hard binary

Bloom

Target

Linear MS-SSIM

JOD 7.699 · PSNR 16.56 · SSIM 0.6242

PU21-MS-SSIM

JOD 7.934 · PSNR 15.04 · SSIM 0.7537

Native ROI · hard binary

Castle

Target

Linear MS-SSIM

JOD 7.535 · PSNR 18.94 · SSIM 0.5815

PU21-MS-SSIM

JOD 8.070 · PSNR 14.96 · SSIM 0.6867

Native ROI · hard binary

Glade

Target

Linear MS-SSIM

JOD 7.588 · PSNR 20.28 · SSIM 0.6327

PU21-MS-SSIM

JOD 8.299 · PSNR 15.74 · SSIM 0.6805

Native ROI · hard binary

Lounge

Target

Linear MS-SSIM

JOD 8.504 · PSNR 31.33 · SSIM 0.8663

PU21-MS-SSIM

JOD 8.719 · PSNR 24.85 · SSIM 0.8483

Native ROI · hard binary

Meridian

Target

Linear MS-SSIM

JOD 7.839 · PSNR 24.89 · SSIM 0.7343

PU21-MS-SSIM

JOD 8.204 · PSNR 20.97 · SSIM 0.7465

Native ROI · hard binary

Rushmore

Target

Linear MS-SSIM

JOD 7.457 · PSNR 22.16 · SSIM 0.6015

PU21-MS-SSIM

JOD 8.027 · PSNR 17.43 · SSIM 0.6435

Native ROI · hard binary

Showgirl

Target

Linear MS-SSIM

JOD 8.034 · PSNR 26.81 · SSIM 0.6905

PU21-MS-SSIM

JOD 8.303 · PSNR 24.64 · SSIM 0.6592

Native ROI · hard binary

Snow

Target

Linear MS-SSIM

JOD 8.821 · PSNR 30.62 · SSIM 0.9050

PU21-MS-SSIM

JOD 8.783 · PSNR 26.16 · SSIM 0.9268

Native ROI · hard binary

Toys

Target

Linear MS-SSIM

JOD 7.707 · PSNR 25.51 · SSIM 0.7604

PU21-MS-SSIM

JOD 8.268 · PSNR 25.62 · SSIM 0.8192

Native ROI · hard binary

Type

Target

Linear MS-SSIM

JOD 7.337 · PSNR 22.90 · SSIM 0.5207

PU21-MS-SSIM

JOD 7.781 · PSNR 19.03 · SSIM 0.4767

Native ROI · hard binary

Water

Target

Linear MS-SSIM

JOD 7.837 · PSNR 20.04 · SSIM 0.6541

PU21-MS-SSIM

JOD 8.322 · PSNR 15.92 · SSIM 0.7616