DEV Community

AI-KSK
AI-KSK

Posted on

FastH3 OmniRef in ComfyUI: 8-Step Distillation, VSA Sparse Acceleration and Multi-Reference Image/Video/Audio Generation

#comfyui #ai #tutorial #videogeneration

FastH3 OmniRef in ComfyUI: 8-Step Distillation, VSA Sparse Acceleration and Multi-Reference Image/Video/Audio Generation

This episode is a mirror release for FastH3 OmniRef, a multi-reference variant of MiniMax H3. The notes below come from the episode subtitles and the published mirror page; no workflow JSON was shipped or parsed, so no node-level parameters are claimed.

What OmniRef changes

The earlier Fast H3 V2 compressed audio-video generation into 8 steps, but mainly for text-to-audio-video. OmniRef strengthens the multi-reference path: the more reference inputs you add, the larger the acceleration gain becomes relative to the original model, which slows down as references grow. It is still slower with more inputs, but the penalty is much smaller.

The official template also added a matching workflow, which is a good signal of how central this variant is.

Measured timings from the episode

  • 1 megapixel, 10-second output, with two reference images plus one video: 7 min 4 s render time.
  • 1 megapixel, 15-second output: 3 min 4 s render time.

These are the author's RTX 5090 measurements on the mirror, not a benchmark across GPUs.

The workflow structure

The quick-start graph and the full graph differ mainly in that the full graph pre-wires several input nodes you can paste into. The multi-material input area has three families:

  1. Images — Load Image for character, costume, prop or scene references. A three-view sheet (front, side, back) works and helps the model understand the character.
  2. Video — Load Video plus a segment selector; a long source can be trimmed to the needed window. The clip's picture and its audio are wired separately: the video's own sound goes to Reference Video Audio.
  3. Standalone audio — Load Audio with start second and duration. Feeding it directly into the multi-reference node uses it as a timbre reference; a short segment is enough. It connects to Reference Audio 0.

After the references, you set the prompt, aspect ratio, megapixel target and duration. The second section loads the H3 OmniRef diffusion model and its encoders.

Deploying the mirror

  1. Open the mirror page: https://www.compshare.cn/images/LtJkEw9QXxfC?referral_code=A7fTQUYY9lCDFJwA6MMvoq
  2. Deploy an instance — this version currently supports RTX 5090 only.
  3. Wait for the image to come up, then open ComfyUI and pick the workflow.

Links

Limits

  • Timings are single-machine observations from the episode, not a controlled benchmark.
  • The mirror currently targets RTX 5090; other GPUs are not claimed to work.
  • Multi-reference quality depends on reference selection and prompt; the mirror is a deployment convenience, not a quality guarantee.

Top comments (0)