Use the native ComfyUI templates
ComfyUI documents support from version 0.30.0. Update ComfyUI, open Template Library → Video, select a MiniMax H3 workflow, and follow the model-download prompts.
- T2V and I2V use the FL2VA model assets.
- R2V uses the separate Ref2VA diffusion weights.
- The documented native canvas uses a 768-pixel short edge and dimensions rounded to a multiple of 32.
Use the upstream serving frameworks
MiniMax lists SGLang, vLLM, and Diffusers as recommended inference frameworks. The repository provides task-specific download scopes, while Diffusers can fetch the components required by its modular pipeline.
- Choose FL2VA for text and first/last-frame workflows.
- Choose Ref2VA for multimodal reference-driven workflows.
- Follow the linked framework recipe because dependencies and commands can change.
Local H3-Base is not the complete hosted 2K system
MiniMax states that H3-Context-IR and H3-Regenerate-2K are not part of the current open-source release. The full 2K validation workflow combines hosted services with a locally deployed H3-Base; a purely local H3-Base workflow validates 768p output.