正在加载视频...
视频加载失败
As promised, FastVideo open-sources MiniMax (official) FastH3 Omni, training code included! 🚀 - Real-time gen @ 12x faster on NVIDIA Blackwell - NVFP4/FP8 for Spark + RTX! - ComfyUI day-0 Collab with NVIDIA AI , Nuva Lab Nuva Lab and Reactor reactor Open source must win! ♥️
20 条评论

@NVIDIAAI @nuvalab @reactorworld Awesome work, can't wait to look at the training code 👀

@MiniMax_AI @NVIDIAAI @nuvalab It's been incredible working with your team!

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld It was a pleasure collaborating with you!

2/7 One request can mix up to 9 images plus video clips and voice references. Name each in the prompt (<Picture 1>, <Video 1>, <Audio 1>) and pick whose face, which motion and whose voice goes where. (video made with FastH3 Omni by @jd_markovchain)

3/7 12x lower end-to-end latency than base MiniMax-H3 Ref2VA at 50 steps. 27 matched workloads on 8x GB200. Same prompts, same references, same 1344x768 output, timed until the MP4 is written. Full input and output examples available on blog!

4/7 How: PDD + DMD distillation. PDD teaches the model to cover several denoising steps per forward pass. DMD pulls the finished output back toward the base model's quality. Training code and reference-aware preprocessing are open. Big shoutout to @NVIDIAAI FastGen team!

5/7 Every gallery clip, same seed and inputs, in 3 versions: Base, OmniRef, and OmniRef + our trained 36-layer NVFP4 (4-bit) VAE decoder. Judge for yourself.

6/7 Serve it: one image + one voice on 4x— GB200, then 8 GPUs across two hosts for bigger reference mixes. Pinned checkpoints and configs. Star the repo, try it in ComfyUI, and tell us what you make. Blog: Model: Uniserve: FastVideo Repo:

7/7 We thank -@nuvalab For the amazing work on data curation, training eval and inference optimizations - @reactorworld For providing the compute and early feedback for production API serving of FastH3 Omni! - NVIDIA FastGen (Julius Berner @julberner, Chao Liu @cxlclb, Arash Vahdat @ArashVahdat ) - NVIDIA Enterprise Products (Pengcheng Li @landy0220, Cliff Woolley) for the collaboration on FastH3! 🚀 Special thank to @NVIDIAAI, @mbzuai, and @vllm_project for the continued support of FastVideo!

@MiniMax_AI @NVIDIAAI @reactorworld 🥳 open source must win. what a journey from ideation to this fasth3 omni milestone!

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld Lfg!!! So epic. Feels like 🎁 thx

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld Weighs:

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld What does a 4090 actually get per second of footage? Real-time on Blackwell is nice but most indie filmmakers aren't renting those.

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld "Open source must win!" Love it guys!

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld thank you so much , u guys r not fake like @fal

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld Incredible work, congrats @haoailab!

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld Who is the current Minimax H3 speed record holder? Last I checked it was around 1s generation time for a 5s video. But this wasn't a Ref model afaik. @grok

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld @ComfyUI you know what to do!

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld Without a 5090, not even NVFP4 can save a 16GB card.

@MiniMax_AI @NVIDIAAI @nuvalab @reactorworld Hey how is this in comparison to trim in turns of speed?


