FastH3 OmniRef Open-Sourced: Up to 11.9x Faster Reference-to-Video, Day-0 ComfyUI Support
Vandy_simp · reddit · 2026-10-21
The FastVideo team at Hao AI Lab has open-sourced FastH3 OmniRef, with ComfyUI support from day one.
- Feed it a prompt plus your own references (images, short clips, a voice) and get a finished video with synced sound
- Just 8 sampling steps instead of 50, up to 11.9x faster than the base MiniMax-H3
- Open weights plus training code; models on Hugging Face, code on GitHub
The team thanked NVIDIA, Nuva Lab, Reactor, and the ComfyUI team, with a full local-deployment blog post dropping the next day.
More from Infra
- Anthropic plans $750B compute spend over 7-10 years; Meta reportedly drives a fifth of revenue — danshipper · 2026-10-21
- NVIDIA ships FP4-Local checkpoints of Qwen3.8-Flash-Next and GLM-5.3-Flash for local AI — NVIDIAAI · 2026-10-21
- Rumor: post-hoc GQA→MLA conversion cuts Qwen 27B KV-cache memory by up to 8.4x — ChrisGPT · 2026-10-21
- EAGLE-4 deep dive: up to 5.21× agentic coding speedup with lossless speculative decoding — zhyncs42 · 2026-10-21
- EAGLE-4 speculative decoding hits 961.6 tokens/s on Qwen3.8-27B, code open-sourced — zhyncs42 · 2026-10-21
- Bittensor miners speed up REAX decision model nearly 3× within one evening of competition — airesearch12 · 2026-10-21