MiniMax H3 Video Lab
You can fill an MBP boot volume with one video DiT pack and still have nothing runnable after unplugging the drive that held the weights. MiniMax H3 is the open-weights omni-modal video model (text / image / video / audio → clip with native stereo audio). In the haus it is a cold pack: occasional creative sessions, stored on external media per the Model Storage Policy.
Lab experiment + scripts: ~/Projects/comfy-lab/experiments/minimax-h3-blender/.
Sanctum-docs heroes now render from this pack (tools/gen_hero_image.py --backend h3 → comfy-lab/scripts/gen-hero-h3.py).
Music 3 is the complementary local song model (ComfyUI 0.33+): ~/Projects/comfy-lab/experiments/minimax-music3/.
Engine matrix note: ~/Projects/blender-lab/ENGINES.md.
Requirements
Section titled “Requirements”| Piece | Gate |
|---|---|
| ComfyUI | ≥ 0.30 core (Desktop app must ship that core; upgrade via brew upgrade --cask comfy) |
| Machine | MBP M4 Max 128 GB (or CUDA box for reliable INT8) |
| Weights | Cold pack on Transcend/Ark, symlinked into ~/Documents/ComfyUI/models/ |
| Road | Bring the external volume (symlinks die if unmounted) |
Weight pack (cold)
Section titled “Weight pack (cold)”On the external volume (example Transcend mount):
ai-models/MiniMax-H3/diffusion_models/— pruned INT8 T2V/I2V/FL2VA and Ref2VA safetensorsai-models/MiniMax-H3/text_encoders/— Heretic Qwen3-VL TE (uncensored conditioning)ai-models/MiniMax-H3/vae/— video + audio VAEs
Sources: Comfy-Org/MiniMax-H3,
ethanfel Heretic TE.
cd ~/Projects/comfy-lab./scripts/download-minimax-h3.sh --ref2va./scripts/link-minimax-h3-models.sh./scripts/minimax-h3-download-status.shStructured pipeline (Blender → H3)
Section titled “Structured pipeline (Blender → H3)”| Step | Tool | Role |
|---|---|---|
| 1 | Blender | Gray clay / greenscreen blockout, 24 fps, clear camera move |
| 2a | ComfyUI Ref2VA | Motion plate as video ref + character/style stills |
| 2b | ComfyUI FL2VA | First/last frames from Blender keyframes |
| 3 | Heretic TE | Uncensored Qwen3-VL-32B conditioning |
| 4 | film-lab | Cut / mix / deliver |
ComfyUI templates
Section titled “ComfyUI templates”After upgrade: Template Library → Video → MiniMax H3 (T2V / I2V / R2V).
Swap the stock text encoder for the Heretic safetensor when you want uncensored embeddings.
Official docs: docs.comfy.org MiniMax H3.
Apple Silicon note
Section titled “Apple Silicon note”Community reports: pruned INT8 may stall on MPS (_int_mm missing). With 128 GB RAM, BF16 or an MLX port is the fallback runtime; CUDA Mini/RunPod runs the INT8 Comfy pack cleanly. Weights on this MBP remain the portable SoT either way.
Smoke verification
Section titled “Smoke verification”# Weights resolvable via ComfyUI model dirsls -la ~/Documents/ComfyUI/models/diffusion_models/minimax_h3_*.safetensorsls -la ~/Documents/ComfyUI/models/text_encoders/*heretic*ls -la ~/Documents/ComfyUI/models/vae/minimax_h3_*
# Lab E2E helper (starts when Comfy listens on :8188)~/Projects/comfy-lab/scripts/e2e-minimax-h3-smoke.shCold pack, hot laptop: keep the ark mounted when you create, and leave the boot volume for things that must survive a reboot alone.