FastH3-Omni in ComfyUI: GGUF vs INT8 setup

Rebel walks through a ComfyUI workflow for FastH3-Omni, a video model he describes as generating synchronized audio from text and ordered media references. He specifies eight sampling steps with guidance set to one, and a limit of 12 references overall: up to nine images, three videos and three audio clips.

For local AI users choosing model files, the tutorial compares INT8 with Rebel's GGUF quantizations. He lists the pruned BF16 file at about 44 GB and INT8 at about 22 GB. His recommendation for INT8 is roughly 24 GB of combined VRAM and system RAM with dynamic VRAM offloading, rather than 24 GB of GPU memory alone. He describes Q4 as a 13 GB option with results close to INT8 but weaker audio, while warning that Q2 and Q3 substantially degrade output.

The setup covers model and text encoder loaders, separate video and audio VAEs, reference images and sampling controls. Rebel also explains his text encoder architecture workaround for the GGUF loader and tells viewers to leave the workflow's Boolean control set to true.

The examples test character consistency, dialogue and anime style transfer. Rebel reports blurred facial details in distant shots, damaged audio in one clip and poor coherence during heavy motion. He favors static cameras and close-ups for cleaner faces, and says animation styles worked best in his tests.