Hao AI Lab's FastH3 Omni Generates Synced Video and Audio 9x Faster

Hao AI Lab’s FastVideo team has released FastH3 Omni, a distilled implementation of MiniMax H3’s omni-reference mode, with native ComfyUI support. Publicly downloadable model files include compressed diffusion checkpoints, text encoders, and video and audio VAEs. ComfyUI confirmed the integration shortly after release, giving developers a standard graph-based workflow for generating video and synchronized sound from image, video, and audio references.

One node, three reference types

FastH3 Omni combines a text prompt with ordered visual and audio references, then generates a single clip with synchronized sound. A workflow can use character stills to preserve appearance, a video to guide motion, and an audio clip to condition speech or other sound.

• Reference limit: Up to 12 inputs in total, subject to caps of nine images, three videos, and three audio clips.
• Required input: Every request needs at least one image or video reference.
• Prompt syntax: References are addressed with tags such as <Picture 1>, <Video 1>, and <Audio 1>.
• Output: Video at 24 FPS with 32 kHz stereo audio.
• Lengths: 124, 243, or 362 frames, corresponding to nominal durations of 5, 10, or 15 seconds.
• Default sampling: Eight steps, guidance 1.0, video shift 12.0, and audio shift 3.0. The shift settings control the denoising schedules for the two modalities.

Eight evaluations replace 50

FastVideo reports that distillation reduces sampling from MiniMax H3’s 50 transformer evaluations to eight. Across the team’s tests on eight Nvidia GB200 GPUs, that produced end-to-end speedups ranging from 5.6 to 11.9 times.