MiniMaxH3-Director-Seed-Hunt Introduction
MiniMaxH3-Director-Seed-Hunt adds a visual directing layer for MiniMax H3 video generation in ComfyUI. Instead of writing one dense prompt and manually tracking every image, video, and audio reference, you can arrange your idea as a timeline of CUTs—short sections of a script—and give each CUT its own direction.
The extension’s main node, MuseMinimaxDirector, helps you:
- Keep characters and settings connected to reference images, video, or audio.
- Turn a longer script into generations short enough for MiniMax H3 to handle.
- Preview different seeds—different starting points for the same generation settings—before choosing a result to refine.
- Inspect the exact prompt sent to the model when a result does not match your intention.
This is a fork of the simpler Muse Minimax Director. Its distinguishing addition is Seed Hunt: you can generate a main result and selectively run up to three more candidate passes. That makes it easier to compare interpretations of a scene before spending more time and computing resources on a higher-resolution refinement.
MiniMax H3 creates the video and audio. This extension organizes your instructions and calls ComfyUI’s existing MiniMax H3 support; it does not contain model weights or replace the model. MiniMax H3’s Community License Agreement applies separately from this extension’s MIT-licensed code.
How MiniMaxH3-Director-Seed-Hunt Works
Think of the Director node as a storyboard assistant. You provide the cast, any visual or audio references, and a sequence of actions. The node turns those pieces into the structured instructions MiniMax H3 expects.
- Build your timeline. Each CUT contains prompt text for part of the video. CUT weights determine how much of the total running time each part occupies. For example, one CUT might describe a character entering a room; the next might describe a close-up as they speak.
- Add references when you need them. In Reference mode, you can supply character images, a location image, reference video, and reference audio. The node assigns tags such as
<Picture 1>and<Audio 1>in the order MiniMax H3 actually receives those references. You do not have to renumber them yourself when some slots are empty. - Let the node assemble the prompt. It organizes Reference-mode instructions into the sections recommended by MiniMax’s official prompt-writing guide, including subjects, shots, sound, and music. The
compiled_promptoutput shows the resulting text. - Render in manageable chunks. A single MiniMax H3 generation is limited to roughly 15 seconds. For a longer video, the node splits the requested duration into chunks and includes a CUT in every chunk its timing overlaps. It also carries continuity cues from one chunk to the next.
- Compare seeds if you choose. The main generation is always
candidate_1. Enablingcandidate_2,candidate_3, orcandidate_4runs additional passes with the same settings but different seeds. You can compare the resulting images and audio, then select a favorite for a separate refinement workflow.
A simple first run: connect the required MiniMax H3 model and encoders, place a character image in Ref 1, write a CUT such as “The character walks along a windswept boardwalk,” and choose a duration. With all extra candidate toggles off, connect images and audio to a Video Combine node. If you enable extra candidates, preview the candidate_N_images and candidate_N_audio outputs instead.
MiniMaxH3-Director-Seed-Hunt Features
Timeline, references, and prompt creation
- CUT-based editor: Write one flowing script as timed segments rather than preparing a separate prompt for every generation call. Adjusting a CUT’s weight changes the portion of the total duration assigned to it. A CUT crossing a chunk boundary is included on both sides.
- Character and location images: Reference mode provides up to nine character image slots and one Location slot. Only filled slots are used. If you fill Ref 1 and Ref 3, they become
<Picture 1>and<Picture 2>—there is no missing<Picture 2>. Location follows the filled character images. - Reference descriptions and prompt structure: Per-slot descriptions help identify what each image represents. The node constructs subject definitions, scene descriptions, retention instructions, shot timing, soundscape, and music sections from the relevant UI fields. You can check the result through
compiled_prompt. - Multi-speaker CUTs: Select speaking characters on a CUT using its speaker chips. With one speaker selected, quoted dialogue can be attributed automatically. With multiple speakers, identify the intended
<Subject N>in your CUT text so the node does not have to guess who speaks. Standalone reference-audio slots pair by position with character slots for voice reference.
Generation modes and continuity
| Mode | What you provide | Best suited to |
|---|---|---|
| Reference | Character, location, video, and/or audio references | Keeping a character or setting recognizable while directing a scene through multiple CUTs |
| First/Last Frame | A starting image in Ref 1 and, optionally, an ending image in Ref 2 | Reaching a specific first frame, last frame, or both |
Reference images guide the model in Reference mode, but do not lock every generated frame to an image. First/Last Frame mode instead locks the supplied images to the corresponding endpoint frames. Its other reference slots are unavailable because this checkpoint does not use Reference mode’s tag system; describe the motion between endpoints in plain CUT text.
For videos spanning several chunks, Reference mode normally carries the previous chunk’s final frame and a short audio tail forward as soft continuity cues. A visible seam can still occur. If you connect the First/Last-Frame checkpoint and enable hybrid_continuation, later chunks use the previous chunk’s final frame as a locked starting frame. This can improve transitions, but those later chunks no longer receive the same direct reinforcement from the original Reference-mode image tags.
Video and audio direction
- Reference video: Use a clip to suggest action, camera movement, or scene structure. You can also describe replacing a person in the clip while retaining its motion and setting. A retention option supports duplicating the video while replacing its character.
- Optional audio from a reference video: Each clip has an “Include this clip’s audio (as its own reference)” toggle, off by default. Enable it when the clip’s speech or timing should influence the result; leave it off when you only want visual motion. For a voice-change or reperformance idea, provide another audio reference for the desired voice and state clearly which audio supplies the words and which supplies the voice.
- Standalone reference audio: Supply a voice sample independently of a video clip when you want to guide dialogue timbre or delivery. Describe the intended use in the CUT.
- Analyze button: The reference-image Analyze control has a gear icon for choosing its vision provider/model. Its character descriptions are designed to avoid mixing in background, pose, or setting details.
Seed Hunt and outputs
candidate_1 is the main run at the chosen seed and costs no additional pass. The other three toggles are independent:
| Toggle | Candidate seed | Additional work |
|---|---|---|
candidate_2 | seed + 1,000,003 | One extra full pass |
candidate_3 | seed + 2,000,006 | One extra full pass |
candidate_4 | seed + 3,000,009 | One extra full pass |
For instance, you can enable only candidate_3 if you want two results total: candidates 1 and 3. Each enabled candidate repeats the generation, so render time and compute use increase with the number enabled. Lowering megapixels can make scouting less costly; choose the most promising result before using the companion Muse Minimax Refine node for a higher-resolution second pass.
The outputs are intentionally different depending on whether you are scouting:
| Output | What to expect |
|---|---|
images and audio | Main frames and audio only when all extra candidate toggles are off. Connect them to a Video Combine node for an ordinary single run. |
candidate_1_images and candidate_1_audio | The main run, available whether or not extra candidates are enabled. |
candidate_2..4_images and candidate_2..4_audio | Results only for the corresponding enabled toggles. Disabled candidate outputs remain empty. |
compiled_prompt | The resolved per-chunk instructions, useful for checking reference tags, dialogue, and continuity text. |
ref_images_used | In Reference mode, the static images actually used for picture tags on the first chunk, in tag order. Connect these to Refine’s ref_images input when using that workflow. It is empty in First/Last Frame mode. |
Empty or blocked outputs stop their connected downstream branches without stopping the rest of the queue. In particular, enabling any extra candidate blocks images and audio by design so an unselected scouting result is not mistaken for your final render.
Other useful controls and included nodes
duration_secondssets the total target length;chunk_duration_secondssets the target size of individual generations. Keep individual chunks within MiniMax H3’s roughly 15-second per-call limit.aspect_ratioandmegapixelscontrol output shape and resolution budget.multiplecontrols resolution rounding;ref_image_sizecontrols the size of reference images sent to H3.resize_methodcontrols how reference and endpoint images fit an output shape that differs from theirs:cropmay remove edges,padretains the image but adds black bars, andstretchcan distort proportions.steps,sampler_name,scheduler,shift_video, andshift_audioexpose generation and sigma-shift controls. Keep them consistent while comparing seeds if you want the seed to be the main variable.MuseModelRoute, the second node included in this extension, sends one MODEL to eitheron_trueoron_false, leaving the other output empty. It can simplify graphs that switch between checkpoints, but its boolean and the Director’smodesetting do not synchronize automatically.
The Director produces frames and audio, not a finished video file. Use a Video Combine node to make a playable output.
MiniMaxH3-Director-Seed-Hunt Models
MiniMax H3 supplies the generation capabilities; the extension does not include its weights. The two checkpoints serve different purposes and are not interchangeable:
| MiniMax H3 checkpoint | Director input | When to use it |
|---|---|---|
| Reference-to-Video | model | Reference mode: guide a scene with character, location, video, and audio references. |
| Image-to-Video (First/Last Frame) | model_fl2va | First/Last Frame mode: lock a supplied starting and/or ending frame. Also needed for Hybrid Continuation in Reference mode. |
You also need a matching CLIP text encoder, video VAE, and audio VAE for the MiniMax H3 workflow. The audio VAE is needed in both modes because the model handles audio and video together internally. The Director handles a separate sigma-shift pass for each checkpoint it uses.
For current model files and guidance, consult the official MiniMax H3 model page. Exact filenames and folder conventions are not listed here because they may change between model and ComfyUI releases. If your ComfyUI version includes a built-in MiniMax H3 workflow template, it can also show the expected model connections.
What's New with MiniMaxH3-Director-Seed-Hunt
v1.2.0
- Independent candidate controls:
candidate_2,candidate_3, andcandidate_4replace the former all-or-nothing Seed Hunt control. You can now pay the generation cost for exactly the extra candidates you want. Turning all three off gives you a normal single run. - Legacy workflow compatibility: The old
seed_huntwidget may remain in saved workflows to preserve widget positions, but it no longer does anything. Enable the new candidate toggles in older graphs that previously relied on it. - New
MuseModelRoutenode: Route a MODEL to one of two outputs using a boolean. The Director can now leave itsmodelinput empty when the selected mode does not need it, making this routing pattern possible. - Updated example workflow: The included scouting graph uses the new candidate toggles and a boolean-based model-routing arrangement.
v1.1.0
- Older-workflow loading fix: A saved graph that produced an error mentioning
Cannot create property 'characters'can now reset its timeline rather than prevent the whole workflow from loading. - More controllable image analysis: The Analyze button gained provider/model settings, and its character descriptions were improved to avoid including pose, background, or setting details.
- New reference-video retention choice: A “duplicate video, replace character” option supports reusing a clip’s scene and motion while changing its subject.
- Revised example workflow: The author selected Patch Sol-Attn as the active speed optimization after same-seed testing. Other tested speed nodes were left bypassed for comparison. These observations describe the example workflow, not a requirement or a guarantee of identical performance on your hardware.
Troubleshooting MiniMaxH3-Director-Seed-Hunt
The images or audio branch does not run, but candidates render.
- Check
candidate_2,candidate_3, andcandidate_4. - If any is on, use the populated
candidate_N_imagesandcandidate_N_audiooutputs for previews. - Turn all three off when you want the ordinary
imagesandaudiooutputs. An empty output in this situation is intentional, not a failed render.
An older workflow’s Seed Hunt checkbox seems to have no effect.
- Look for the new
candidate_2,candidate_3, andcandidate_4controls. - Enable whichever additional passes you want.
- Do not rely on the saved
seed_huntwidget; it is retained only for older workflow compatibility.
A saved workflow reports Cannot create property 'characters' or loses its timeline.
- Use a version that includes the v1.1.0 loading fix.
- If the old node still gives you trouble, delete that Director node and add a fresh one.
- Re-enter its references and CUTs. Replacing the node is the author’s recommended safest remedy for an affected saved workflow.
ComfyUI reports No module named 'comfy_extras.nodes_minimax_h3'.
- Check whether your ComfyUI version includes the stock MiniMax H3 nodes.
- Update ComfyUI core, not just this extension.
- Restart ComfyUI and reopen the workflow.
Reference video or audio cannot be decoded or trimmed. The extension uses the Python package av for uploaded clips. Check that it is available in the Python environment your ComfyUI installation uses.
The wrong checkpoint runs, or the node fails after switching modes.
- Check the Director’s
mode: Reference or First/Last Frame. - Confirm the corresponding checkpoint reaches
modelormodel_fl2va. - If you use
MuseModelRoute, set its boolean consistently with the Director’smode; changing one does not change the other. - For Hybrid Continuation, confirm the First/Last-Frame checkpoint is available through
model_fl2va.
A character, voice, or location is not following the intended reference.
- View
compiled_promptthrough a text-preview node. - Check the resolved
<Picture N>,<Video N>, and<Audio N>tags rather than assuming they match the numbers of partly filled UI slots. - Check each reference description and make the desired relationship explicit in the CUT text.
- For multi-speaker dialogue, select the speakers and identify the relevant
<Subject N>when more than one character speaks.
A long video has a visible jump between chunks. Check chunk_duration_seconds and the CUTs around that boundary. Reference mode carries continuity cues, but they are not a hard frame lock. If a matching First/Last-Frame checkpoint is available, try hybrid_continuation; remember that later chunks then trade some direct image-reference reinforcement for a locked transition.
The included example workflow shows missing nodes. Open workflows/muse_minimax_h3_director_scout_v1.json as a guide to the scouting and refinement layout. It contains optional preview, memory-management, upscaling, and speed-related nodes that the Director itself does not require. Replace or bypass optional nodes if you do not need them. For example, its decorative ModelPreviewOverrideKJ wrapper can be removed and rewired straight through, and its iToolsPreviewText can be replaced with another STRING preview node.
Frequently asked questions
- Do I have to generate four candidates? No. Candidate 1 always runs; enable zero to three additional candidates independently.
- Does Seed Hunt automatically choose the best result? No. Preview the candidates and choose the one that best fits your creative goal. The linked Refine workflow can then perform a second pass on your selection.
- Does lowering
megapixelsmake a later high-resolution result automatic? No. It changes the resolution budget of the scouting run. A higher-resolution second pass uses a separate refinement workflow. - Can First/Last Frame mode use all nine character references? No. It repurposes Ref 1 and Ref 2 as endpoint frames and does not use Reference mode’s picture, video, or audio tag system.
- Does the node output a ready-to-play video? No. Connect its populated frame and audio outputs to a Video Combine node.
Learn More about MiniMaxH3-Director-Seed-Hunt
- Included Seed Hunt example workflow: Explore candidate previews, model routing, prompt inspection, and the companion refinement branches. Its reference slots are deliberately empty so you can supply your own material.
- MiniMax H3 official model page: Check the source model and its current documentation.
- MiniMax H3 official prompt-writing guide: Learn more about reference tags, shot descriptions, and the prompt format behind the Director’s compiled text.
- Muse Minimax Refine: Read about the companion node used for a selected candidate’s second-pass refinement.
- Original Muse Minimax Director: Compare the simpler single-generation version with this Seed Hunt fork.
- MiniMax H3 Community License Agreement, extension LICENSE, and NOTICE: Review the separate terms and attribution information before using or redistributing work built around the model.
