Muse Minimax Director

If you've ever hand-built a MiniMax H3 prompt in ComfyUI, you know the pain this node exists to kill. Stock H3 wants numbered <Picture N> / <Video N> / <Audio N> tags that have to line up exactly with the order you feed references, a strict six-section prompt format, and - here's the real killer - a hard ceiling of roughly 15 seconds per single generation call. Muse Minimax Director (from Muse Collective's Seed Hunt pack) sits on top of ComfyUI's stock MiniMaxH3ReferenceToVideo and MiniMaxH3ImageToVideo nodes and does all that bookkeeping for you. You write a script broken into CUTs on a visual timeline; it builds the prompt, numbers the tags, and chunks anything too long into separate H3 calls with continuity carried between them.

The timeline UI has three parts: up to 9 character reference slots (plus a Location background), a reference video slot with a per-clip audio toggle, and reference audio for voice cloning. CUTs are timed segments with their own prompt text. That's the same shape as the LTX Director phenomenon - a model whose native interface can't express "this, then that, at this moment," so the community wraps a timeline around it. Where it gets clever is the bookkeeping underneath: reference items are always numbered the way H3 itself assigns them (iteration order, not slot position), and each CUT's text is included in every chunk it actually overlaps in time, not a fixed per-chunk split.

Two generation modes matter. Reference (Omni) uses up to 9 images, 3 videos, and 3 audio references via soft conditioning. First/Last Frame hard-locks your first two slots to frame 0 and the final frame using the separate model_fl2va checkpoint. There's also hybrid_continuation, which routes chunk boundaries through the First/Last Frame checkpoint for a hard pixel-lock instead of soft carry-over - a real trade, since continuation chunks then lose fresh character-reference reinforcement.

The inputs a beginner actually sets: model (the Reference-to-Video checkpoint), clip, vae, and audio_vae - the tooltip is right, that audio VAE is required in both modes because H3 always builds a joint audio+video latent. Then duration_seconds and chunk_duration_seconds (keep chunks under H3's ~15s trained range), seed, and steps. Don't sleep on compiled_prompt - the single best debugging tool on the node, showing exactly what got sent to H3 including every resolved tag. images/audio wire into a Video Combine; ref_images_used feeds the companion Muse Minimax Refine node.

Seed Hunt is the reason this fork exists. Instead of paying full price to discover your render ignored the prompt, you scout cheap: tick candidate_2/candidate_3/candidate_4, each runs one extra pass at a fixed seed offset (+1,000,003, etc.) and fills its own output pair. Watch the trap: when any candidate toggle is on, the main images/audio outputs are deliberately blocked (ExecutionBlocker) - they only mean "the one real result" when all three toggles are off. Wire the candidate_N outputs instead while scouting. The old single seed_hunt checkbox is dead - it still renders in saved workflows but does nothing.

Install

ComfyUI Manager, search "Muse Minimax Director". Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/MiniMaxH3-Director-Seed-Hunt

Then restart, and pip install av - it's used to decode and trim reference video/audio. The H3 weights aren't in the repo; grab MiniMaxAI/MiniMax-H3 from HuggingFace yourself (two checkpoints if you want First/Last Frame mode), plus CLIP, video VAE, and audio VAE.

The elephant: the license

Before you get excited: the MiniMax H3 Community License excludes the US, EU, UK, and South Korea from its territory - including outputs, not just weights. If you're in one of those regions, running these weights locally is unlicensed on the face of the text. The node author flags this plainly in the README; it's not a ComfyUI problem, it's a model problem, but it's the first thing to check.

Troubleshooting

• ModuleNotFoundError: No module named 'comfy_extras.nodes_minimax_h3' - your ComfyUI core predates native H3 support. Update ComfyUI core itself, not this node.
• H3 is heavy. ~42.5GB full weights and no verified consumer-VRAM floor at release; a 3060 generation is measured in geological timescales. Scout at low megapixels, refine at full res.
• Cannot create property 'characters' on load - old saved workflow crash; delete the node and add a fresh one.
• This is a control layer only. Credit MiniMax H3 in anything you ship - the README asks for it, and the license's NOTICE requirements are real for redistribution.