The useful answer

  • Select FL2VA or Ref2VA before building the graph.
  • Normalize and label reference media outside the prompt.
  • Test short outputs before scaling resolution or duration.
  • Validate both the video stream and native stereo audio.

Choose the graph family

A ComfyUI H3 workflow begins with the conditioning task. Use FL2VA for text generation or a defined first and last frame. Use Ref2VA when the prompt depends on ordered reference images, clips or audio.

Starting from the wrong checkpoint often looks like a prompt-quality problem. Resolve the task-model match before tuning guidance or sampling settings.

Build the graph in clear stages

Keep the workflow readable: load shared components, prepare multimodal inputs, encode conditioning, run the H3 sampler, decode visual and audio latents, then mux the streams into the final file. Group and label each stage so a broken reference does not become a mystery downstream.

  • Inputs: prompt plus optional frames, references or audio.
  • Conditioning: preserve the intended order and role of each asset.
  • Generation: keep a baseline seed and settings for comparison.
  • Output: inspect duration, frame rate, stereo audio and container playback.

Use references deliberately

Reference-heavy prompts work best when each input has a clear job: subject identity, scene layout, motion, pacing or sound. Remove redundant media and describe relationships directly. More inputs do not automatically produce more control.

Normalize source orientation, duration and audio levels before the graph. This makes failures easier to isolate and reduces accidental conflicts between the prompt and media.

Create a review loop

Save the prompt, seed, checkpoint, adapter and graph with each approved output. Review picture and sound together; native audio is part of the generation, not an optional afterthought. Use short draft passes to settle composition and a controlled delivery pass for the chosen shot.

Questions answered

What people ask about comfyui minimax h3

Which checkpoint should a ComfyUI MiniMax H3 workflow use?

Use FL2VA for text or first/last-frame tasks and Ref2VA for ordered reference-media tasks.

Does the workflow output audio?

Yes. H3 generates native stereo audio, and the workflow should decode and mux it with the video stream.

How should I troubleshoot a workflow?

Return to a short baseline graph, confirm model paths and versions, and add references or optimizations one at a time.

Primary reading

Sources behind this guide

MiniMax H3 official repositoryReference implementation, checkpoints and prompt guidance.MiniMax H3 integration directoryOfficially maintained map of runtimes, training tools and community projects.MiniMax H3 official model cardWeights, architecture, capabilities and license links.