AR Video Synthesis on VBench++ Long 1-minute
98.58Subject ConsistencyRolling Sink
Evaluation Results
| Method | Links | |||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Rolling SinkTraining duration=5s at 16 FPS2026.02 | 98.58 | 96.94 | 63.08 | 69.68 | 100 | 69.98 | 80.23 | 100 | 21.59 | 25.03 | 23.16 | 78 | 98.16 | 98.65 | 74.69 | 18.91 | 1.375 | |
| Self ForcingTraining duration=5s at 16 FPS2026.02 | 96.79 | 96.53 | 59.16 | 69.8 | 86.8 | 36.39 | 64.33 | 71.21 | 10.79 | 22.2 | 19.91 | 68.86 | 97.63 | 98.14 | 48.57 | 20.99 | 2.4375 | |
| LongLiveTraining duration=5s at 16 FPS, LoRA usage=w/o LoRA2026.02 | 96.68 | 95.88 | 58.5 | 65.19 | 97.8 | 58.02 | 77.12 | 96.83 | 25.4 | 23.98 | 21.6 | 88.57 | 96.43 | 97.3 | 75.92 | 20.18 | 2.1875 |