FastH3 Omni Full BF16
video model conversion for ComfyUI
1 min readupdated
This repository provides a complete .safetensors weights file for running FastH3 Omni 8-Step V1 in ComfyUI. The model generates video from text prompts and reference inputs along with synchronized audio. It was converted from the original shards while preserving the full AdaLN matrices and all transformer blocks. This release is useful for ComfyUI users who want to run new multimodal video generators locally.
Key features
- Full export of all 50 transformer blocks and 32 PDD heads
- Single-file
.safetensorsformat ready for ComfyUI - Reference input support for generation control
- Video and synchronized audio generation in 8 steps
What you can do
- Generate video with synchronized audio from text prompts and references in ComfyUI
- Use official workflows for FastH3 Omni
- Pair the model with compatible text encoders and VAEs from MiniMax H3
Example
Place the model file in ComfyUI/models/diffusion_models/, load the official video workflow template, and configure the text encoders along with the audio and video VAEs.
Read more
DetailsFile size: ~70.19 GB
Precision: BF16 with FP32 output heads
Parameters: 8 steps, guidance 1.0
Requirements: high VRAM and system RAM requirements, depending on video resolution and length














Comments