FastH3 Omni in ComfyUI: Which Files to Download and How References Work
ComfyUI Wiki explains how to run FastH3 Omni with the built-in video_fasth3_ominiref_r2v template, which model files to place and which build to pick for your GPU. It also shows how the prompt points at image, video and audio references by position.
Key points
- Place one diffusion model, the Qwen3-VL text encoder, the video VAE and the audio VAE; no custom node pack is needed
- The int8 diffusion file (22.2 GB) matches the template default; BF16 is 44.2 GB, and FP8 (23.1 GB) and NVFP4 (13.8 GB) builds are separate repos
- Use NVFP4 on RTX 5090 and DGX Spark, FP8 with layerwise offload on RTX 4090
- Template defaults: 8 steps, guidance 1.0, sigma shift 12 for video and 3 for audio, 1344×768, 124 frames
- References are numbered by type and position in the prompt; at least one image or video is required
- Reference sizing "match" is faster, "max" keeps a 2048 px short edge for identity at several times the cost