Speed Any mode 4 steps
FastH3 VSA 4-step (experimental, fastest)
FastVideo's 4-step VSA distill with NVIDIA (90% sparsity); biggest gains on long sequences, still an experimental path (live arena rank on its LoRA page).
How to run it
- 1 Grab Kijai's converted FastH3 VSA int8 weights (see sources)
- 2 Wire the VSA nodes, or route attention through core Block Sparse Attention (SLA/VSA modes)
- 3 Sample with 4 steps; start at 736×416–960×544 for short clips
The official validation used multi-GPU + fastvideo-kernel; on a single consumer card, spot-check quality.
Representative GPUs
864 × 480 · 8s · 32GB RAM / fast-disk; click a GPU for its full matrix.
| GPU | VRAM | Status | Per clip | VRAM peak |
|---|---|---|---|---|
| RTX 4060 8GB | 8GB | Slow | 5 min 4 s | 7G |
| RTX 3060 12GB | 12GB | Slow | 4 min 49 s | 11G |
| RTX 4070 12GB | 12GB | Slow | 2 min 51 s | 11G |
| RTX 4070 Ti Super 16GB | 16GB | OK | 2 min 6 s | 15G |
| RTX 4080 Super 16GB | 16GB | OK | 1 min 53 s | 15G |
| RTX 4090 24GB | 24GB | OK | 1 min 24 s | 23G |
| RTX 5090 32GB | 32GB | OK | 1 min 2 s | 28G |
| RTX PRO 6000 Blackwell 96GB | 96GB | OK | 59 s | 49G |
1344×768 · 5s: 48 of 49 listed GPUs workable, from about 6GB VRAM.
FAQ
What's the minimum GPU for this recipe?
About 6GB+ of VRAM for 1344×768 · 5s (48 of 49 listed GPUs workable). Short 736×416 clips run on less.
How does this recipe combine with other acceleration?
Step distillation picks one option; attention backends (Sage/Kitchen/SLA/VSA) and caches stack on top. See the conflicts table for exact rules.