MiniMax H3/Updated 2026-11-03
MiniMax H3 ComfyUI
This page is a practical MiniMax H3 ComfyUI checklist, not a rewrite of docs.comfy.org. It focuses on the traps that waste a download: wrong Comfy version, mixing BF16 HF weights with Comfy-Org templates, and resolution steps that blow the H3-Base pixel-area limit.
What you actually install
- ComfyUI templates use pruned/quantized files from Comfy-Org/MiniMax-H3. MiniMax reviewed Comfy-Org revision
4cc1d817in their local-deploy guide. - Official HF BF16 checkpoints (MiniMaxAI/MiniMax-H3) are the FL2VA / Ref2VA dumps for SGLang, vLLM, and diffusers. Do not mix benchmark claims between the two paths.
- Open H3-Base does not include H3-Context-IR or H3-Regenerate-2K. Local ComfyUI output at a 768 short edge is not the full hosted 2K pipeline.
Setup steps
- 1. Update ComfyUI to 0.30.0 or later. H3 templates ship with ComfyUI 0.30+. Older builds will not show the MiniMax H3 entries in Template Library > Video.
- 2. Open a built-in H3 template. Choose MiniMax H3 T2V, I2V, or R2V. These templates pull pruned/quantized files from Comfy-Org/MiniMax-H3 — different artifacts from the official BF16 Hugging Face dump.
- 3. Download the required models from the scan prompt. Restart ComfyUI if the model list does not refresh. Keep FL2VA and Ref2VA files in the folders the template notes specify.
- 4. Set resolution inside the H3-Base area limit. MiniMax’s local-deploy guide reviewed template revision 3c1df78: use 0.98 megapixels with multiples of 32, or set 1344×768 for 16:9. Avoid the 1.0 megapixel step (1376×768), which can exceed the 768×1344 area limit.
- 5. Queue and check for an MP4 with stereo audio. A healthy run returns 24 FPS video plus synchronized stereo audio in one file. If you only get silent video, the audio VAE path or template wiring failed.
Workflow downloads from Comfy Org’s day-0 post: T2V / I2V / R2V templates. Prefer those JSON files over screenshots of node graphs.
VRAM: what vendors claim vs what we measured
Comfy Org’s day-0 post says modulation pruning plus int8 convrot quantization cut the footprint from about 123.6 GB full precision to about 42.5 GB on the smallest variants, and that dynamic offloading can run on an RTX 3060-class card. MiniMax’s own ComfyUI baseline table says published minimum GPU memory and performance figures are not listed — validate on your hardware. Spacebunny has not published an independent VRAM measurement on this page.
Errors that usually mean a setup mismatch
- Template missing — ComfyUI older than 0.30.0, or Template Library not refreshed after update.
- OOM mid-run — resolution above the H3-Base area limit, or mixing full BF16 weights into a template built for Comfy-Org quant files.
- Silent MP4 — audio VAE / stereo path not wired; re-open the official template instead of rebuilding from memory.
- Identity drift in R2V — too few references or a prompt that never names which picture owns the face/wardrobe. Start from the R2V template and the prompt library on MiniMax H3 prompts.
Related MiniMax H3 pages
spacebunny.me is independent and not affiliated with MiniMax, ComfyUI, fal.ai, or Hugging Face. This hub does not generate video on our servers.