FastVideo/FastH3-Omni-FP8-Comfy

FastH3 Omni FP8 for ComfyUI

The ComfyUI-native FP8 diffusion model for FastH3 Omni. It generates video with synchronized 32-kHz stereo audio from a prompt and ordered image, video and audio references in eight sampling steps.

Release collection · BF16 release · Shared ComfyUI components

Install

Place fasth3_omniref_fp8_comfy.safetensors in:

ComfyUI/models/diffusion_models/fasth3_omniref_fp8_comfy.safetensors

Download a compatible text encoder, video VAE and audio VAE from FastH3 Omni Comfy and place them in ComfyUI/models/text_encoders/ and ComfyUI/models/vae/. The audio VAE and the text encoder's reference-vision support are required. Select a text-encoder format supported by your GPU.

FP8 uses ComfyUI’s native quantization layout and kernels. Quantized output can differ from BF16. This is the ComfyUI format; use the separate FastVideo-format model for FastVideo inference.

Generate video and audio

Use a current ComfyUI build with native MiniMax-H3 support and the MiniMax H3 Reference to Video conditioning node. Supply matching text-encoder, video-VAE and audio-VAE components.

• Eight sampling steps, guidance 1.0.
• Video shift 12.0 and audio shift 3.0.
• Reference inputs use <Picture 1>, <Video 1> and <Audio 1> tags in the prompt.
• The audio VAE is required for synchronized audio output.

ComfyUI uses its own model layout, sampler and runtime kernels; outputs can differ from FastVideo and UniServe.

Other runtimes

• UniServe — GB200/B200, RTX PRO 6000 Blackwell and Hopper: use the BF16 release and deployment guide.
• FastVideo local AI — DGX Spark and RTX 5090: use the FastVideo-format NVFP4 QAD model and local inference guide.

Local AI resources

Run FastH3 Omni on DGX Spark or RTX 5090 through FastVideo using the FastVideo-format NVFP4 QAD weights. Start with the FastVideo cookbook and FastH3 Omni recipe builder; use the Spark recipe or RTX 5090 recipe and the local inference guide for the AdaLN table, offload and decoder setup.

FP8 weights are a separate FastVideo variant; the FP8 recipe for RTX 4090 illustrates the non-FP4 path with layerwise offload and compatible BF16 encoder components. Use each runtime's matching checkpoint format.

Acknowledgements

FastH3 Omni was built in close collaboration with Nuva Lab, which co-developed the model with FastVideo, defined and curated the omni reference dataset, and brought production-grounded creative video workflows to FastH3. For dedicated production deployment, fine-tuning on your own data and a creative agent on top, see Nuva Lab's FastH3-Omni page.

Thanks also to the NVIDIA FastGen team.

License

Use of these weights is governed by the MiniMax-H3 Community License Agreement. Bundled ComfyUI tooling retains its own licenses in conversion/upstream/.