• 5-second video of people, dogs, and cats running on the beach
• 832x480, 8 steps, total time taken 142 seconds (sampling 104 seconds + decoding 38 seconds)
• SPARK GB10 128GB VRAM + INT8/NVFP4 quantization
15-second video complete! A woman in a red scarf in a meadow + glowing butterfly storm + aurora scene.
Performance comparison:
| | 5 seconds (124 frames) | 15 seconds (362 frames) |
| ----------------------------------------------------------------------- | ----------- | ------------ |
| Steps per frame | 12 seconds | 63 seconds |
| Sampling | 104 seconds | 502 seconds |
| Decoding | 38 seconds | 93 seconds |
| Total time | 2 minutes 30 seconds | ~10 minutes |
| Although the number of frames is tripled, the steps per frame are 5 times slower because the attention matrix becomes squared.
The quality is better than I thought, and the speed is fast.
This is scary good ㄷㄷㄷㄷㄷ
