Super-resolution ×4 — SeedVR2 vs Real-ESRGAN (Set14, RTX 5090)

Timings & VRAM

Variant	CFG	Steps	Load	Avg gen / img	Peak VRAM	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓	NIQE ↓	MUSIQ ↑	CLIP-IQA ↑	Success
Bicubic x4 (baseline)	—	—	—	0.0s	—	24.06	0.6688	0.4425	0.2240	7.63	32.42	0.4938	14 / 14
Ground truth (original HR)	—	—	—	—	—	—	—	—	—	4.94°	66.94°	0.7418°	14 / 14
Real-ESRGAN x4 (fl16)	—	1	3.9s	0.1s	0.25 GB	22.48	0.6220	0.2460	0.1647	4.75	65.09	0.6247	14 / 14
SeedVR2 3B (bf16, 1-step)	1.0	1	7.6s	0.7s	8.11 GB	21.39	0.5664	0.2293	0.1475	3.90	71.99	0.7565	14 / 14
SeedVR2 7B (bf16, 1-step)	1.0	1	9.2s	0.9s	17.70 GB	22.06	0.6030	0.1877	0.1342	3.97	72.25	0.7804	14 / 14
SeedVR2 7B-sharp (bf16, 1-step)	1.0	1	9.2s	0.8s	17.70 GB	21.81	0.5927	0.1919	0.1300	3.83	72.34	0.8074	14 / 14

Shaded columns are no-reference — they judge “does this look like a real photograph?” with no ground truth. They have no winner: the ° ground-truth row is the calibration line, and out-scoring it means inventing texture prettier than the original. The unshaded columns are full-reference and measure distance to the ground truth. A model can win one axis and lose the other — that is the perception–distortion tradeoff, not a contradiction.

Drag the slider to compare A vs B; switch either dropdown to pick any variant. The number next to each dropdown is that variant's generation time for this example. Click the input thumbnail to view it full size.