HDR Hologram Loss Sweep · Progress Report

Three-scene average · final hard reconstruction

Best value in each metric is green

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓
L1	7.390	19.65	0.4090	0.014563
L2	7.542	18.31	0.3669	0.017754
Charbonnier	7.392	19.65	0.4095	0.014545
Direct CVVDP	8.326	14.46	0.2865	0.037681
MDSI	5.334	12.18	0.1509	0.062606
HaarPSI	5.534	12.24	0.1628	0.061599
DISTS	5.257	12.26	0.1489	0.061531
TOPIQ-FR	5.204	12.11	0.1477	0.063641
CVVDP luma heatmap x Charb	7.445	19.73	0.3998	0.014440

HDR scene · run 260722_041913

Anime2

Candidate vs L2 JOD +0.209 PSNR +2.53 SSIM +0.0710

Final metrics at 300 iterations

Hard reconstruction is primary; soft metrics are muted

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	7.874	23.62	0.4792	0.004347	7.909	24.61	0.5091
L2	7.843	21.31	0.4000	0.007403	7.897	21.73	0.4158
Charbonnier	7.876	23.63	0.4800	0.004332	7.912	24.63	0.5100
Direct CVVDP	8.674	16.24	0.2871	0.023751	8.707	16.27	0.2922
MDSI	5.662	13.79	0.1552	0.041743	5.665	13.78	0.1550
HaarPSI	5.742	13.78	0.1594	0.041837	5.750	13.77	0.1590
DISTS	5.546	13.90	0.1522	0.040759	5.546	13.89	0.1520
TOPIQ-FR	5.476	13.70	0.1499	0.042650	5.474	13.70	0.1501
CVVDP luma heatmap x Charb	8.052	23.84	0.4710	0.004134	8.105	24.91	0.5007

Convergence curves

Evaluated every 20 iterations

Visual comparison

All cards use the same scene and state

TargetReference

L1Hard · iter 100

L2Hard · iter 100

CharbonnierHard · iter 100

Direct CVVDPHard · iter 100

MDSIHard · iter 100

HaarPSIHard · iter 100

DISTSHard · iter 100

TOPIQ-FRHard · iter 100

Anime2 CVVDP luma heatmap x Charb reconstruction

CVVDP luma heatmap x CharbHard · iter 100

HDR scene · run 260722_043842

Bloom

Candidate vs L2 JOD -0.267 PSNR +0.46 SSIM +0.0128

Final metrics at 300 iterations

Hard reconstruction is primary; soft metrics are muted

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	6.870	15.33	0.4032	0.029332	6.946	15.72	0.4191
L2	7.123	14.88	0.3875	0.032486	7.153	15.09	0.3934
Charbonnier	6.873	15.33	0.4036	0.029308	6.948	15.73	0.4195
Direct CVVDP	8.235	12.78	0.3573	0.052693	8.252	12.79	0.3623
MDSI	4.881	11.65	0.1740	0.068389	4.885	11.65	0.1738
HaarPSI	5.346	11.77	0.1868	0.066479	5.360	11.77	0.1856
DISTS	4.861	11.67	0.1716	0.068036	4.863	11.67	0.1714
TOPIQ-FR	4.877	11.59	0.1709	0.069349	4.876	11.59	0.1710
CVVDP luma heatmap x Charb	6.856	15.34	0.4003	0.029243	6.929	15.74	0.4164

Convergence curves

Evaluated every 20 iterations

Visual comparison

All cards use the same scene and state

TargetReference

L1Hard · iter 100

L2Hard · iter 100

CharbonnierHard · iter 100

Direct CVVDPHard · iter 100

MDSIHard · iter 100

HaarPSIHard · iter 100

DISTSHard · iter 100

TOPIQ-FRHard · iter 100

Bloom CVVDP luma heatmap x Charb reconstruction

CVVDP luma heatmap x CharbHard · iter 100

HDR scene · run 260722_052941

Castle

Candidate vs L2 JOD -0.233 PSNR +1.28 SSIM +0.0148

Final metrics at 300 iterations

Hard reconstruction is primary; soft metrics are muted

Loss	JOD ↑	PSNR ↑	SSIM ↑	L2 ↓	Soft JOD	Soft PSNR	Soft SSIM
L1	7.427	20.00	0.3445	0.010010	7.464	20.73	0.3631
L2	7.661	18.74	0.3133	0.013373	7.778	19.22	0.3302
Charbonnier	7.428	20.00	0.3449	0.009996	7.466	20.74	0.3636
Direct CVVDP	8.068	14.37	0.2151	0.036599	8.096	14.35	0.2219
MDSI	5.459	11.10	0.1234	0.077687	5.457	11.09	0.1235
HaarPSI	5.513	11.16	0.1423	0.076480	5.517	11.15	0.1431
DISTS	5.365	11.20	0.1228	0.075799	5.363	11.19	0.1227
TOPIQ-FR	5.259	11.03	0.1223	0.078925	5.260	11.02	0.1225
CVVDP luma heatmap x Charb	7.428	20.02	0.3281	0.009943	7.466	20.79	0.3460

Convergence curves

Evaluated every 20 iterations

Visual comparison

All cards use the same scene and state

TargetReference

L1Hard · iter 100

L2Hard · iter 100

CharbonnierHard · iter 100

Direct CVVDPHard · iter 100

MDSIHard · iter 100

HaarPSIHard · iter 100

DISTSHard · iter 100

TOPIQ-FRHard · iter 100

Castle CVVDP luma heatmap x Charb reconstruction

CVVDP luma heatmap x CharbHard · iter 100