Minimax H3 Video

MiniMax H3 (Hailuo) — local video

This skill teaches the local-weights MiniMax H3 path in ComfyUI. It is the pilot for #1155 (Official vs Empirical sources) because MiniMax publishes a real prompting guide. Cite that guide by URL. Do not copy it into this repo.

Two products, two cost models — pick one

They share a brand and must not be mixed.

API nodes do not take MiniMaxH3SigmaShift or Sage-attention patches. Local nodes do not spend API credits. If the user asked for Hailuo cloud, stop and use the API nodes + their key; do not download 40 GB of weights.

MiniMaxH3Director is a third-party pack (muse-collective-26/MiniMaxH3-Director), not core. Do not require it for T2V / I2V / R2V.

License — cite, do not copy

Local weights and MiniMax's own documentation sit under the MiniMax H3 Community License. Materials includes the Documentation. The agreement's Applicable Territory excludes the United States, the EU, the UK, and South Korea. This skill does not reproduce MiniMax's skills/h3-prompt-writing/ SKILL.md or the prompting-guide prose. Linking to a public URL is the #1155 requirement.

This is not legal advice. Tell a US/EU/UK/KR user that the local path is territory-restricted and that the paid API is a separate product under MiniMax platform terms.

Prefer the Comfy-Org template over hand-wiring

ComfyUI ≥ 0.30.0 (templates in the 0.33 line). These are core comfyui-workflow-templates graphs in the frontend Template Library → Video, not installer packs and not custom-node example_workflows:

list_packs action:"list_templates" will not list them. enqueue_workflow action:"run_template" will not resolve video_minimax_h3_t2v / _i2v / _r2v. That action only loads bundled installer packs, and there is no packs/minimax-h3-* yet. Do not call it until a pack exists. panel_load_workflow needs pack:, a disk path:, or an inline UI graph. A Template Library basename is none of those.

Load path that works:

1. Preferred. Ask the user to open Template Library → Video → MiniMax H3: Text to Video (or Image to Video / Reference to Video). Pick the local video_minimax_h3_* cards, not the api_minimax_h3_* paid partner templates.
2. Agent, no UI click. Fetch the UI JSON from https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_t2v.json (or _i2v / _r2v; raw.githubusercontent.com is the same files), save it with save_workflow action:"save" filename:"video_minimax_h3_t2v.json", then panel_load_workflow path:"video_minimax_h3_t2v.json". Same pattern as video-extend (stage on disk, then path:). Do not pass the GitHub URL as path: or pack:.

After it lands, retarget the subgraph's exposed widgets (prompt, duration, turbo_mode, megapixels). Official T2V/I2V graphs wrap MiniMaxH3ImageToVideo inside a subgraph (type is a UUID). Do not flatten that interior unless you are hand-building.

Hand-building the subgraph is slower and easy to get wrong.

Comfy tutorial (wiring, not MiniMax's prompt formula): https://docs.comfy.org/tutorials/video/minimax/minimax-h3

Models (Comfy-Org INT8 pack)

All from huggingface.co/Comfy-Org/MiniMax-H3. Download with download_model action:"download".

Turbo LoRAs (4–8 steps instead of ~20)

The Comfy-Org T2V template already switches these on with turbo_mode:

Kijai conversions live at Kijai/MiniMax-H3_comfy (loras/) and experimental W4A8 at Kijai/MiniMax-H3-experimental. Same job (low-step / low-VRAM). Prefer the Comfy-Org / lightx2v filenames the template already names; only switch to a Kijai file if that is what is on disk.

4-step is faster and softer; 6 to 8 steps is the usual sharpness compromise.

Output spec

Duration → frame length (Comfy-Org template math, 17-frame blocks):

max(5, round(seconds * 24)) + (5 - (max(5, round(seconds * 24)) % 17)) % 17

That is the 17k+5 grid. Do not invent a WAN-style 4n+1 length.

Node graph (local T2V / I2V)

From the Comfy-Org T2V subgraph (core nodes, not the Markdown notes):

ResolutionSelector (aspect, megapixels, multiple=32) → width, height

UNETLoader (fl2va int8)
  ├─ LoraLoaderModelOnly (turbo LoRA) ─┐
  └────────────────────────────────────┤ ComfySwitchNode (turbo_mode)
                                       ▼
                         BasicGuider + BasicScheduler + KSamplerSelect(res_multistep)
                                       ▼
CLIPLoader (type=minimax, qwen3vl 32b) → MiniMaxH3ImageToVideo
VAELoader (video vae) ─────────────────→   prompt, width, height, length
optional first_frame / last_frame ─────→   → CONDITIONING + LATENT
                                       ▼
                         SamplerCustomAdvanced → LATENT
                                       ├─ VAEDecode (video vae) → IMAGE
                                       └─ VAEDecodeAudio (audio vae) → AUDIO
                                       ▼
                         CreateVideo (fps=24) → SaveVideo

MiniMaxH3ImageToVideo is T2V when both image sockets are empty, I2V with first_frame, FL2VA with both frames. Do not add a second T2V-only node. The manga-director-codex MiniMax H3 prompt adapter (prompt_adapters/minimax_h3.json) declares that as mode text_to_video alongside I2V / FL2VA / L2VA / R2V (#2786).

R2V replaces the UNet with ref2va and the conditioner with MiniMaxH3ReferenceToVideo. Do not load fl2va into an R2V graph.

Local-only helpers

Sage roughly doubles speed. H3 has mixed dtypes, so console lines about falling back to pytorch attention on some layers are expected.

Sampler defaults (Comfy-Org template)

Guider is BasicGuider (CFG-distilled checkpoint, so do not crank CFG). Seed via RandomNoise.