Super-resolution ×2 — SeedVR2 vs Real-ESRGAN (DIV2K valid, RTX 5090)

Timings & VRAM

Variant	CFG	Steps	Load	Avg gen / img	Peak VRAM	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓	NIQE ↓	MUSIQ ↑	CLIP-IQA ↑	Success
Bicubic x2 (baseline)	—	—	—	0.0s	—	32.08	0.9126	0.1692	0.0918	4.76	52.52	0.5463	14 / 14
Ground truth (original HR)	—	—	—	—	—	—	—	—	—	2.93°	66.75°	0.6450°	14 / 14
Real-ESRGAN x2 (fp16)	—	1	4.0s	0.2s	0.62 GB	27.98	0.8268	0.1534	0.0758	3.67	67.70	0.5945	14 / 14
SeedVR2 3B (bf16, 1-step)	1.0	1	7.8s	3.2s	12.69 GB	23.07	0.6789	0.2455	0.1321	2.86	70.04	0.7045	14 / 14
SeedVR2 7B (bf16, 1-step)	1.0	1	9.1s	3.3s	22.09 GB	25.22	0.7439	0.1657	0.0879	2.91	68.68	0.6671	14 / 14
SeedVR2 7B-sharp (bf16, 1-step)	1.0	1	9.0s	3.3s	22.09 GB	26.24	0.7849	0.1600	0.0781	2.98	67.14	0.6314	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.