Video SR ×4 — SeedVR2 vs FlashVSR (VideoLQ, RTX 5090)

Timings, VRAM & quality

Variant	CFG	Steps	Load	Avg gen / clip	Peak VRAM	NIQE ↓	MUSIQ ↑	CLIP-IQA ↑	Flicker ↓	Success
Bicubic ×4 (baseline)	—	—	—	0.1s	—	8.12	24.66	0.3097	0.0876	11 / 11
FlashVSR v1.1 full (bf16, Wan2.1 VAE)	1.0	1	5.5s	73.1s	13.97 GB	3.84	57.79	0.4650	0.0998	11 / 11
FlashVSR v1.1 tiny (bf16, TCDecoder)	1.0	1	5.5s	25.0s	13.13 GB	4.02	52.82	0.4002	0.0958	11 / 11
SeedVR2 3B (bf16, 1-step)	1.0	1	8.0s	71.5s	25.90 GB	4.95	46.83	0.2959	0.1002	11 / 11
SeedVR2 7B (fp8 DiT, 1-step)	1.0	1	9.3s	74.8s	26.01 GB	5.65	41.51	0.2527	0.0995	11 / 11
SeedVR2 7B-sharp (fp8 DiT, 1-step)	1.0	1	10.2s	76.0s	26.01 GB	5.40	42.26	0.2531	0.1003	11 / 11

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 divider to diff A vs B; the two clips are frame-synchronised. Switch either dropdown to pick any variant. Flicker and temporal seams are invisible in a still and invisible at 1× — watch at 0.25× and at 2–4×.

The videos on this page are H.264, and therefore lossy — they exist so you can watch them. Every metric in the table was computed on the raw frames, before encoding. For reference, a crf-18 round-trip of one model's own output measured PSNR 38.6 dB / max abs error 59 / 255. Judge fine texture from the 1:1 stills, not from the video.