FastH3-Omni-Comfy: Simultaneous video and audio generation using image, video, and audio references

A ComfyUI-exclusive diffusion model based on MiniMax H3. It takes a text prompt along with aligned image, video, and audio reference materials as input to generate video and synchronized audio. Diffusion model files are provided in two formats, BF16 and INT8 ConvRot, allowing selection based on hardware specifications.

It must be used with the Qwen3VL 32B text encoder and a dedicated video/audio VAE. Reference materials are specified using <Picture 1>, <Video 1>, and <Audio 1> tags within the prompt, and it operates with 8-step sampling and a guidance setting of 1.0. NVFP4 acceleration is supported only on Blackwell hardware; other devices must use BF16 or INT8.

This model, co-developed with Nuva Lab, assumes a ComfyUI environment with native MiniMax-H3 support. Because it uses ComfyUI's proprietary runtime kernels, output results may differ from those of FastVideo or UniServe. For local environments such as DGX Spark or RTX 5090, it is recommended to separately utilize the NVFP4 QAD model in FastVideo format.

Usage is governed by the MiniMax-H3 community license terms. It is suitable for developers looking to perform reference-based video generation in ComfyUI workflows. It records over 260,000 monthly downloads, with related Spaces and derivative models being actively shared.