HDR Native-ROI Loss Sweep

Four-scene mean

Final hard-binarized reconstruction at iteration 300.

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓
L1	8.763	25.18	0.6889	0.003436
L2 (amplitude)	8.709	23.83	0.6236	0.004486
Charbonnier	8.768	25.19	0.6897	0.003427
CVVDP direct	8.950	14.67	0.3282	0.035279
MDSI	5.641	11.42	0.1461	0.073056
HaarPSI	5.560	11.54	0.1541	0.071527
DISTS	5.327	11.74	0.1252	0.067900
TOPIQ-FR	5.173	11.17	0.1202	0.077398
CVVDP luma heatmap × Charbonnier	8.932	25.09	0.6809	0.003491
LPIPS-VGG (official)	5.645	11.68	0.2669	0.068826

Native ROI · run 260722_205702

Anime2

Candidate − L2JOD +0.326PSNR +1.53SSIM +0.0678

Final metrics

Hard reconstruction at iteration 300; soft values shown for diagnosis.

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	8.098	27.70	0.7198	0.001698	8.230	30.70	0.7926
L2 (amplitude)	8.448	26.18	0.6537	0.002411	8.774	29.67	0.7254
Charbonnier	8.100	27.72	0.7209	0.001690	8.229	30.72	0.7936
CVVDP direct	9.041	16.25	0.3010	0.023696	9.151	16.38	0.3107
MDSI	5.789	12.46	0.1378	0.056763	5.804	12.45	0.1377
HaarPSI	5.705	12.71	0.1458	0.053518	5.712	12.72	0.1455
DISTS	5.450	12.79	0.1206	0.052568	5.450	12.78	0.1201
TOPIQ-FR	5.312	12.14	0.1144	0.061135	5.309	12.13	0.1141
CVVDP luma heatmap × Charbonnier	8.774	27.71	0.7215	0.001695	8.967	30.80	0.7933
LPIPS-VGG (official)	5.849	12.66	0.2512	0.054161	5.863	12.66	0.2505

Convergence

Hard and soft metrics sampled every 20 iterations.

Visual comparison

Switch every reconstruction together.

TargetReference

L1Hard · 100

L2 (amplitude)Hard · 100

CharbonnierHard · 100

CVVDP directHard · 100

MDSIHard · 100

HaarPSIHard · 100

DISTSHard · 100

TOPIQ-FRHard · 100

CVVDP luma heatmap × CharbonnierHard · 100

LPIPS-VGG (official)Hard · 100

Native ROI · run 260722_205702

Bloom

Candidate − L2JOD +0.240PSNR +0.65SSIM +0.0438

Final metrics

Hard reconstruction at iteration 300; soft values shown for diagnosis.

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	9.080	22.23	0.6999	0.005984	9.315	25.25	0.7474
L2 (amplitude)	8.873	21.54	0.6536	0.007009	9.198	24.53	0.7090
Charbonnier	9.080	22.24	0.7005	0.005976	9.317	25.25	0.7479
CVVDP direct	8.944	13.09	0.4136	0.049066	9.046	13.12	0.4286
MDSI	5.333	10.94	0.1370	0.080606	5.341	10.96	0.1364
HaarPSI	5.368	10.49	0.1068	0.089286	5.381	10.44	0.1058
DISTS	5.038	11.10	0.1162	0.077698	5.036	11.09	0.1152
TOPIQ-FR	4.953	10.58	0.1033	0.087522	4.950	10.57	0.1031
CVVDP luma heatmap × Charbonnier	9.113	22.19	0.6974	0.006046	9.358	25.19	0.7441
LPIPS-VGG (official)	5.586	11.40	0.3068	0.072502	5.596	11.42	0.3044

Convergence

Hard and soft metrics sampled every 20 iterations.

Visual comparison

Switch every reconstruction together.

TargetReference

L1Hard · 100

L2 (amplitude)Hard · 100

CharbonnierHard · 100

CVVDP directHard · 100

MDSIHard · 100

HaarPSIHard · 100

DISTSHard · 100

TOPIQ-FRHard · 100

CVVDP luma heatmap × CharbonnierHard · 100

LPIPS-VGG (official)Hard · 100

Native ROI · run 260722_211807

Castle

Candidate − L2JOD +0.224PSNR +1.65SSIM +0.0682

Final metrics

Hard reconstruction at iteration 300; soft values shown for diagnosis.

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	9.044	26.71	0.6975	0.002134	9.202	31.54	0.7725
L2 (amplitude)	8.772	24.71	0.6018	0.003378	8.879	29.15	0.6848
Charbonnier	9.049	26.73	0.6987	0.002124	9.205	31.56	0.7737
CVVDP direct	8.801	14.45	0.2540	0.035868	8.889	14.58	0.2674
MDSI	5.881	10.62	0.1334	0.086693	5.886	10.61	0.1337
HaarPSI	5.671	10.99	0.1669	0.079529	5.674	10.99	0.1685
DISTS	5.488	11.10	0.1100	0.077713	5.483	11.10	0.1095
TOPIQ-FR	5.241	10.42	0.1097	0.090857	5.234	10.39	0.1095
CVVDP luma heatmap × Charbonnier	8.996	26.36	0.6700	0.002313	9.141	31.03	0.7469
LPIPS-VGG (official)	5.575	10.73	0.1753	0.084464	5.567	10.71	0.1728

Convergence

Hard and soft metrics sampled every 20 iterations.

Visual comparison

Switch every reconstruction together.

TargetReference

L1Hard · 100

L2 (amplitude)Hard · 100

CharbonnierHard · 100

CVVDP directHard · 100

MDSIHard · 100

HaarPSIHard · 100

DISTSHard · 100

TOPIQ-FRHard · 100

CVVDP luma heatmap × CharbonnierHard · 100

LPIPS-VGG (official)Hard · 100

Native ROI · run 260722_211805

Glade

Candidate − L2JOD +0.100PSNR +1.20SSIM +0.0494

Final metrics

Hard reconstruction at iteration 300; soft values shown for diagnosis.

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	8.831	24.06	0.6383	0.003928	8.975	27.41	0.7110
L2 (amplitude)	8.745	22.88	0.5851	0.005147	8.895	25.60	0.6554
Charbonnier	8.841	24.07	0.6388	0.003920	8.981	27.42	0.7117
CVVDP direct	9.016	14.88	0.3443	0.032485	9.081	15.00	0.3518
MDSI	5.562	11.66	0.1761	0.068162	5.566	11.64	0.1759
HaarPSI	5.496	11.95	0.1970	0.063773	5.502	11.95	0.1981
DISTS	5.333	11.96	0.1540	0.063619	5.326	11.96	0.1529
TOPIQ-FR	5.184	11.54	0.1533	0.070078	5.178	11.53	0.1528
CVVDP luma heatmap × Charbonnier	8.845	24.08	0.6345	0.003908	8.976	27.44	0.7071
LPIPS-VGG (official)	5.570	11.93	0.3343	0.064178	5.573	11.93	0.3334

Convergence

Hard and soft metrics sampled every 20 iterations.

Visual comparison

Switch every reconstruction together.

TargetReference

L1Hard · 100

L2 (amplitude)Hard · 100

CharbonnierHard · 100

CVVDP directHard · 100

MDSIHard · 100

HaarPSIHard · 100

DISTSHard · 100

TOPIQ-FRHard · 100

CVVDP luma heatmap × CharbonnierHard · 100

LPIPS-VGG (official)Hard · 100