LastKV-Wan-14B
480p causal text-to-video generation with LastKV, based on Wan2.1-T2V-14B.
FP32 generator weights from step 10,000. This is the EMA version; raw weights are also available.
Paper results
VBench scores (%, higher is better), from Table 8 of the paper.
| Model | Quality | Semantic | Total |
|---|---|---|---|
| Transformer | 82.2 | 72.7 | 80.3 |
| LastKV | 83.0 | 76.4 | 81.6 |
These results correspond to the step 10,000 EMA checkpoint.
Full VBench breakdown (16 metrics)Individual scores from the same paper EMA evaluation (%, higher is better).
| Metric | Transformer | LastKV |
|---|---|---|
| Subject consistency | 93.4 | 94.6 |
| Background consistency | 95.3 | 95.1 |
| Temporal flickering | 96.6 | 97.8 |
| Motion smoothness | 98.4 | 98.9 |
| Dynamic degree | 68.1 | 73.3 |
| Aesthetic quality | 61.1 | 61.6 |
| Imaging quality | 66.9 | 62.8 |
| Object class | 86.2 | 92.1 |
| Multiple objects | 66.3 | 75.5 |
| Human action | 95.0 | 96.0 |
| Color | 80.3 | 86.5 |
| Spatial relationship | 68.3 | 71.7 |
| Scene | 48.9 | 47.2 |
| Appearance style | 20.3 | 21.6 |
| Temporal style | 22.2 | 23.0 |
| Overall consistency | 24.4 | 25.4 |
Use the matching LastKV implementation with config.json and the Wan2.1 VAE/text encoder. Load all 12 weight shards through the safetensors index.