FastVideo releases FastWan2.1-T2V-1.3B, a 3-step distilled video model hitting 16 FPS on one H100
AI video model releases
FastVideo releases FastWan2.1-T2V-1.3B, a 3-step distilled video model hitting 16 FPS on one H100
FastWan2.1-T2V-1.3B-Diffusers is finetuned from Wan-AI's Wan2.1-T2V-1.3B using a new Sparse-distill strategy that combines DMD distillation with VSA (Video Sparse Attention) in one training pass.
The model supports 3-step inference and generates 61x448x832 resolution video at up to 16 FPS on a single H100 GPU.
Training used 600k synthetic latents from the FastVideo 480P Synthetic Wan dataset, run on 4 nodes with 32 H200 GPUs, global batch size 64, for 4000 steps (about 12 hours).
VSA attention sparsity was set to 0.8 during training, and the model can generate at any resolution though quality degrades outside the trained 61x448x832 setting.
Inference and finetuning scripts are open sourced in the FastVideo GitHub repo, with support ranging from H100 down to consumer 4090 GPUs and Mac.
cho.sh