GitHub radar
Make a video with synced audio from a prompt and clips
FastH3 Omni packages MiniMax H3's video model for ComfyUI, the app that runs these models. One run gets you a finished video with matching audio from a prompt and references.
It's an AI model: from a text prompt plus references — an image, a video, an audio clip — it generates a new video with matching audio. Files include a diffusion model, a text encoder, a video VAE (builds the video) and an audio VAE (builds the audio, same vae/ folder) — offered as a light INT8 version for weak cards and a more precise BF16 version (more accurate but heavier) for strong cards.
When it helps
Useful when you want a finished clip with audio in one pass, e.g. for content work. Skip it if your ComfyUI build has no MiniMax H3 support — this isn't plug-and-play.
Pros
- One file instead of assembling components separately
- Audio comes out already synced to the video
- A light version exists for weaker cards
Cons
- Needs a GPU with 24GB+ VRAM — weak cards won't manage
- Needs MiniMax H3 support in ComfyUI plus a reference node
- Commercial use follows MiniMax's own license, not open source
How to set it up — step by step
- 1Open your AI agent (Claude Code, Codex) in your ComfyUI project folder.
- 2Send the agent the text from the block below.
- 3Have the agent pick the light or precise weight version for your GPU.
- 4Have the agent gather references: an image and an audio clip.
- 5Check the test clip with audio that comes out.
- 6Check MiniMax's license before any paid commercial use.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the FastH3 Omni build based on MiniMax H3 by FastVideo in ComfyUI for me: https://huggingface.co/FastVideo/FastH3-Omni-Comfy — pick the lighter variant for my GPU, gather references (an image and a piece of audio), and make a test clip with sound.
You need a GPU with at least 24 GB of VRAM — with the audio codec and the text model included, the total weight of the files is more than 40 GB, so this is a job for a strong GPU, not for a regular laptop.
Open on Hugging Face▌ More finds