# Muse Minimax Director (Seed Hunt) **Timeline-based director node for MiniMax H3 in ComfyUI, with Seed Hunt scouting** Built by [Muse Collective](https://musecollective.co.uk) — write a single flowing script broken into CUTs, drop in reference character images/video/audio, and let the node handle chunking, prompt-per-chunk splitting, and reference-tag numbering for you. This is a fork of [Muse Minimax Director](https://github.com/muse-collective-26/MiniMaxH3-Director) that adds a **Seed Hunt** toggle — scout 4 candidate seeds in one run at low resolution, then pick the best one and refine it at full resolution with the companion [Muse Minimax Refine](https://github.com/muse-collective-26/Muse-MiniMax-H3-Refine) node. The original repo stays as the simpler, single-generation version; this one is for anyone who wants the scouting workflow built in. ![ComfyUI Custom Node](https://img.shields.io/badge/ComfyUI-Custom%20Node-orange?style=flat-square) ![MiniMax H3](https://img.shields.io/badge/MiniMax-H3-blue?style=flat-square) ![License](https://img.shields.io/badge/license-MIT-green?style=flat-square) --- ## Changelog ### v1.2.0 - **Seed Hunt is now three independent toggles instead of one all-or-nothing switch.** `candidate_2` / `candidate_3` / `candidate_4` each run their own extra pass (or don't) — pick exactly which ones you want to pay for instead of always paying for all 3. `candidate_1` still always runs for free, same as before. Turning all three off runs exactly like a normal single-generation workflow, same as `seed_hunt` off used to. The old `seed_hunt` widget still exists on already-saved workflows but no longer does anything — it's kept only so its saved widget position doesn't shift and corrupt older saved workflows' stored data, not because it's still functional. If your old workflow relied on it, tick the three new toggles instead. - **New node: `MuseModelRoute`**, bundled in this same repo — the reverse of a switch. Feed it one MODEL and a boolean, and it sends that model to exactly one of two outputs (`on_true` / `on_false`), leaving the other empty. Pairs with this node's `model`/`model_fl2va` inputs so a single boolean can pick which checkpoint (Reference or First/Last-Frame) is actually active, instead of maintaining two separate model chains permanently wired in. See [Switching between Reference and First/Last Frame checkpoints, in detail](#switching-between-reference-and-firstlast-frame-checkpoints-in-detail). - **`model` no longer forces a checkpoint load it doesn't need.** Previously, the sigma-shift step run on `model` happened unconditionally regardless of `mode` — meaning `model` had to always be a real, connected checkpoint even in First/Last Frame mode using `model_fl2va`. It's now skipped when `model` is empty, which is what makes the `MuseModelRoute` pattern above possible without crashing. Doesn't change anything if you wire `model` normally, as before. - **Updated example workflow** — now wired with `MuseModelRoute` and a single boolean switch for the Reference/First-Last-Frame model choice, and the three independent candidate toggles in place of the old single `seed_hunt` checkbox. ### v1.1.0 - **Fixed a crash on old saved workflows.** If you'd saved a workflow before this update and it failed to load with an error mentioning `Cannot create property 'characters'`, that's fixed — the node now resets its timeline gracefully instead of blocking the whole workflow from loading. If you hit this, the safest fix is to delete the Muse Minimax Director node from your old workflow and add a fresh one in its place, then re-enter your references/CUTs. - **Analyze button improvements:** it no longer leaks background/pose/setting details into character descriptions, and it now has its own settings (gear icon) so you can pick which vision provider/model it uses instead of being locked to a hidden default. - **Reference video:** added a "duplicate video, replace character" retention option for people who want to reuse a reference video's motion/scene but swap in a different character. - **Updated example workflow** (`workflows/muse_minimax_h3_director_scout_v1.json`) — now wired with **Patch Sol-Attn** as the active speed optimization. Based on direct same-seed testing, Sol-Attn was the only speed node that gave a real, consistent speed-up without any visible quality loss on hands/hair. The other speed nodes (Sage Attention variants, EasyCache, Spectrum) are left in the workflow but bypassed, so you can switch them on to compare for yourself — in our tests they either made no measurable difference or came with a quality cost. --- ## Powered by MiniMax H3 This node is a control layer built entirely on top of **MiniMax H3**, an open-weights omni-modal video generation model published by [MiniMax](https://www.minimax.io/). MiniMax H3 is the actual model doing every bit of the generation work here — this repository contains none of the model's weights, training code, or inference architecture. All of that lives in MiniMax H3 itself and in ComfyUI core's own stock support for it (`comfy_extras/nodes_minimax_h3.py`). What this node adds on top is a visual timeline/scripting layer, automatic chunking for videos longer than a single H3 generation, and correct handling of H3's reference-tag numbering — nothing more. - **Model page:** [huggingface.co/MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) - **License (MiniMax H3 Community License Agreement):** [huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE) - **Official prompt-writing guide:** [VIDEO_PROMPT_WRITING_GUIDE_ref_en.md](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md) You must download MiniMax H3's weights yourself, directly from the official source above, and you are bound by MiniMax's own license terms for using them — this node does not change, replace, or grant any rights over that license. See [Credits & Licensing](#credits--licensing) below for the full detail, and see [NOTICE](NOTICE) and [LICENSE](LICENSE) in this repository. **If you build something with this node, please credit MiniMax H3 in your own output/description — "Powered by MiniMax H3" — the same courtesy this README extends to them.** --- ## What it does MiniMax H3 is a strong omni-modal model, but its native inputs are low-level: numbered `

Name	Type	Default	Description
model	MODEL		—
switch	BOOLEAN	false	—

Name	Type	Description
on_true	MODEL	—
on_false	MODEL	—

Muse Model Route

Inputs (2)

Outputs (2)