Robotics Image-to-Video Generation on PAI-Bench-G
89.6Grasp SuccessCosmos-Predict-2.5 + Predict-and-Perturb
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Cosmos-Predict-2.5 + Predict-and-PerturbNFE=57, Time Overhead=1.6×2026.01 | 89.6 | 76.3 | 75.1 | |
| Wan2.2-I2V-A14B + Predict-and-PerturbNFE=60, Time Overhead=1.5×2026.01 | 85.7 | 80.3 | 75.5 | |
| Cosmos-Predict-2.5 + Verifier (best-of-4)NFE=140, Time Overhead=4.0×2026.01 | 84.4 | 72.3 | 75.3 | |
| Wan2.2-I2V-A14B + NFE×2NFE=80, Time Overhead=2.0×2026.01 | 83.1 | 76.7 | 75.5 | |
| Wan2.2-I2V-A14B + Verifier (best-of-4)NFE=144, Time Overhead=4.0×2026.01 | 80.5 | 78.1 | 75.3 | |
| Cosmos-Predict-2.5NFE=352026.01 | 79.2 | 71.7 | 75.1 | |
| Cosmos-Predict-2.5 + NFE×2NFE=70, Time Overhead=2.0×2026.01 | 78.6 | 72.6 | 75.1 | |
| Wan2.2-I2V-A14BNFE=402026.01 | 77.3 | 77.4 | 75.3 |