MiniMax H3 ComfyUI Guide — Workflows, Models, VRAM, and Online Use
Run MiniMax H3 in ComfyUI with official text, image, and reference workflows, understand model files and hardware tradeoffs, or use H3 online.
MiniMax H3 has official ComfyUI workflows for text-to-video, image-to-video, and reference-to-video. Local H3 gives creators control over model files and workflow graphs, but it also requires substantial downloads, system memory, storage, and render time. Cloud generation removes that setup in exchange for usage-based credits.
This guide separates the official workflow from rapidly changing community optimizations so you can choose the right path.
Start with the official ComfyUI templates
Update ComfyUI or ComfyUI Desktop before troubleshooting custom nodes. Open the workflow template browser, select Video, then choose the MiniMax H3 workflow that matches your input:
- Text to Video: starts from a written audiovisual prompt.
- Image to Video: animates a first frame and can guide toward a last frame.
- Reference to Video: combines identity, motion, visual, video, or audio references.
The official templates define the required graph and link to the compatible files. Starting from the template is safer than importing an old screenshot or an unofficial workflow built before H3 support landed.
Understand the model file groups
A local workflow needs more than one diffusion-model file. The current ComfyUI package separates the H3 video model, a Qwen-based text encoder, video VAE, and audio VAE. Reference and first/last-frame workflows may use different H3 model variants.
Keep each file in the folder shown by the official template. A common failure is placing the audio VAE where the video VAE belongs, or installing a model variant that does not match the selected workflow.
After downloading files, restart ComfyUI if selectors remain empty. Open every model selector and verify its value rather than assuming the template found a similarly named file.
Local hardware: distinguish minimum from practical
ComfyUI reports that its smallest optimized H3 package reduces the model memory footprint substantially and can use dynamic VRAM offloading. That does not mean every resolution or duration will be fast on a 12 GB GPU.
Community results illustrate the range:
- one RTX 4090 Laptop test reported a five-second 960×540 clip in roughly three minutes with a pruned INT8 model;
- an RTX 4060 Ti reference-to-video test reported much longer generation time for demanding clips;
- users report that RAM, SSD speed, resolution, step count, attention method, and whether models are reloaded all change total time.
These are individual configurations, not guaranteed benchmarks. Record your own resolution, steps, model variant, VRAM, RAM, and end-to-end time before comparing performance.
A practical local setup checklist
- Update to the latest stable ComfyUI build.
- Load an official H3 template before adding custom nodes.
- Download the exact model, text encoder, video VAE, and audio VAE requested by that template.
- Confirm each file is in the expected ComfyUI model directory.
- Restart and verify every selector resolves to an installed file.
- Begin at a modest resolution and short duration.
- Save a known-good workflow before testing attention or caching changes.
- Change one performance setting at a time and compare the complete output with sound.
Common MiniMax H3 ComfyUI errors
Missing model or empty selector
The file is absent, in the wrong directory, or ComfyUI has not refreshed its model list. Use the links inside the official template and restart after the download completes.
Video VAE and audio VAE errors
H3 produces both picture and sound, so the workflow contains separate decoding paths. Confirm the video VAE and audio VAE are not swapped.
Out-of-memory during sampling or decoding
Reduce resolution, duration, and concurrent work. Close other GPU-heavy apps and confirm offloading is active before changing several optimizations at once. System RAM and fast local storage matter when weights move between GPU, CPU, and disk.
Soft, distorted, or unstable output
Return to the stock workflow. Community accelerators can trade fidelity for speed, and a setting that helps one GPU may hurt another. Compare against a baseline with the same prompt, seed, resolution, and steps.
Reference mode ignores an asset
Explain the role of every image, video, and audio reference in the prompt. Reference mode is not a guarantee that all sources will influence every frame. Remove unnecessary inputs and test the strongest reference first.
ComfyUI or online MiniMax H3?
| Choose local ComfyUI when | Choose online H3 when |
|---|---|
| You want graph-level control | You want to start without model downloads |
| You already own suitable hardware | You need occasional or predictable usage |
| Long local render time is acceptable | You value managed infrastructure |
| You want to test community optimizations | You want Text, Image, and Reference in one UI |
| You can manage storage and updates | You want private account history and downloads |
Local is not truly zero-cost: hardware, electricity, storage, maintenance, and time still matter. Cloud generation is not automatically cheaper either. Use the path that matches your volume and tolerance for setup.
Prompting is shared across both paths
The same creative principles apply locally and online: define the opening, observable action, camera path, soundscape, and final composition. Reference mode still benefits from explicit roles and preservation rules.
Use the MiniMax H3 prompt guide before benchmarking. For capability tradeoffs, read the MiniMax H3 review, or use MiniMax H3 online without configuring a local graph.