Scored v1.0 mini profile: 100 tasks, one sample each. The gallery follows the final 170-task training/evaluation scope based on benchmark v1.1; its 70 newly added tasks do not yet have published metric files and are therefore not folded into this table. Every scored generation is resampled to its target's resolution, frame rate and frame count; PSNR is the benchmark's official score (rgb-psnr-v1: per-pixel MSE over all frames in [0,1], equal-weight mean within a task, macro-mean over the 100 scored tasks). The other columns are computed on the same resampled pair. n counts the scored samples a model returned. The copy-input row scores the input video itself against the target.
| Rank | Model | n | PSNR ↑ |
|---|---|---|---|
| — | Copy-input baseline | 100 | 20.11 |
| 1 | Kling O3 Standard | 100 | 16.76 |
| 2 | Gemini Omni Flash 1.0 | 88 | 15.99 |
| 3 | Grok Imagine (extend) | 100 | 14.10 |
| 4 | Seedance 2.5 | 90 | 13.65 |
| 5 | LTX-2.3 Extend | 100 | 13.50 |
| 6 | HY-OmniWeaving | 99 | 13.11 |
| 7 | LTX-2.3 (IC-LoRA v2v) | 100 | 12.65 |
| 8 | Wan-VACE 14B | 100 | 12.02 |
| 9 | MiniMax H3 | 100 | 11.24 |
| 10 | Cosmos3 Nano | 100 | 8.51 |
| 11 | Cosmos3 Super | 100 | 8.49 |
Per-sample scores: model_means.csv · each model's per_sample.csv under /media/scores/<model>/.