Motion-controllable video generation on Motion Controllability Benchmark 206-clip
0.44Latency (s)AR-Drag
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| AR-DragBase Model=Wan2.1-1.3B-I2V, Denoising steps=3-step, Hardware=8 NVIDIA H20 GPUs2025.10 | 0.44 | 28.98 | 187.49 | 4.07 | 99.48 | 4.37 | |
| Self-ForcingBase Model=Wan2.1-1.3B-I2V, Fine-tuned=true2025.10 | 0.95 | 34.47 | 315.87 | 3.7 | 99.2 | 4.06 | |
| DragAnything2025.10 | 68.76 | 38.13 | 367.74 | 3.22 | 98.11 | 3.63 | |
| DragNUWAArchitecture=DiT-based2025.10 | 94.26 | 36.31 | 376.39 | 3.3 | 97.59 | 3.71 | |
| ToraArchitecture=DiT-based2025.10 | 176.51 | 32.84 | 283.43 | 3.86 | 98.55 | 3.97 | |
| MagicMotionParameters=5B2025.10 | 1,426.37 | 30.04 | 230.53 | 4.01 | 98.71 | 3.95 |