Muse MiniMax Director V1.4

MiniMax H3 is a genuinely strong open-weights video model - 33B, omni-modal, with audio generated in the same pass as the picture. But the stock ComfyUI path is fiddly in three ways: you hand-order its <Picture N> / <Video N> reference tags, each generation call tops out around 15 seconds, and anything longer than that means stitching chunks where the cut loves to visibly jump. Muse MiniMax Director V1.4 is a control layer that turns all three into UI. You write your video as a script of timed CUTs on a visual timeline, drop references into slots, and it handles the prompt compilation, chunking, tag numbering, and continuity for you. No API, no key - it's a local layer over ComfyUI's own stock H3 nodes.

The mechanism is the interesting part. Each CUT is a timed script segment the node maps into every H3 chunk it overlaps in time (straddling CUTs land in both), then compiles each chunk into MiniMax's full six-section reference prompt - subject_definitions, summary, retention_analysis, and so on - with tags numbered by H3's actual assignment rule (iteration order, not the numeric slot you dropped the image into). That numbering fix alone saves real headaches.

A mode toggle picks which checkpoint drives the run: Reference (Omni) takes up to seven image references plus three video and three audio clips for character/scene consistency; First/Last Frame hard-locks your start and end frames positionally instead (no reference-tag system there, so you write plain motion prose); Hybrid does both. Two more headline features: two-stage sampling (on by default - a few low-res steps, then a direct latent upscale to full resolution on the same noise schedule) and Seed Hunt, which runs your timeline at up to four low-res seeds so you can pick a winner before spending real compute.

The inputs you'll actually set

• model - the Reference/Omni (ref2va) H3 checkpoint. Wire the First/Last-Frame checkpoint into model_fl2va for First/Last Frame or Hybrid mode.
• clip, vae, and audio_vae - all required. The audio VAE isn't optional even in First/Last Frame mode, because H3 always builds a joint audio+video latent internally.
• duration_seconds and chunk_duration_seconds - your total length and the per-call target (keep it under H3's ~15s ceiling; the final chunk absorbs whatever's left over).
• Seed Hunt: enable_seed_hunt plus candidate_count (1–4), and two_stage_seed_hunt_latent_only if you want each scout to stop after the cheap Stage 1.

Everything else - sampler, scheduler, and the continuity toggles that hard-freeze the previous chunk's real ending into the next chunk's opening so seams stop looking like seams - has sane defaults. Leave them alone until something looks wrong.

Outputs and where they go

images + audio feed a standard Video Combine node to get a playable file. compiled_prompt is your debug window - it shows exactly what got sent to H3, including every resolved tag. ref_images_used, the candidate_1..4_latent outputs, and ref_audio_1..3_used are what you feed the bundled Muse Minimax Refine companion for the high-resolution second pass on a scouted winner. One trap baked into the design: with Seed Hunt on, images/audio are deliberately blocked - a scouting run isn't a picked result, so read your candidates instead.

Install

ComfyUI Manager: search Muse Minimax Director and click Install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/MiniMaxH3-Director-V1.2

Restart ComfyUI. Manager installs the pack's requirements.txt (faster-whisper, used for the in-timeline audio transcription that builds timed shot breaks). PyAV (av) is imported at module top level, so a node that won't register usually just needs pip install av. Two more installs the README underplays: multi-chunk continuity needs ComfyUI-H3-Motion-Context-MultiRef, and because two-stage sampling is on by default the code hard-requires the Comfyui_Minimax_h3_latent_Upscaler pack plus a trained upscale checkpoint in ComfyUI/models/latent_upscale_models/ - the dropdown shows a placeholder until you add one, and the node errors until you do.

Then the model itself, which isn't bundled: grab the MiniMax H3 weights from Hugging Face (MiniMaxAI/MiniMax-H3) - both checkpoints, plus CLIP and the video/audio VAEs - and read the H3 Community License before you do. It's permissive-ish but geofenced: users in the US, EU, UK, and Korea aren't licensed to run the open weights at all, a detail that keeps biting people.

Where people get burned

If two-stage sampling errors at runtime, read the message - it literally tells you which pack or model file is missing. A ModuleNotFoundError: No module named 'comfy_extras.nodes_minimax_h3' means your ComfyUI core predates native H3 support, so update core, not this node. And if a multi-chunk render still cuts oddly, check compiled_prompt first, then make sure the pack is current - the changelog shows real fixes for reference videos being capped at 200 frames and played at the wrong frame rate. This repo is obsessively documented - the changelog reads like a bug diary, which is the energy you want from something sitting between you and a 15-minute render.