Nodes/MiniMaxH3-Director-Seed-Hunt/Muse Minimax Director
ComfyUI Node

Muse Minimax Director

A ComfyUI node in Muse Collective with 25 inputs and 12 outputs.

By muse-collective-26·Created 19 days ago·Updated 7 days ago· 22
Muse Minimax Director
  • model
  • clip
  • vae
  • audio_vae
  • model_fl2va
  • images
  • audio
  • compiled_prompt
  • ref_images_used
  • candidate_1_images
  • candidate_1_audio
  • candidate_2_images
  • candidate_2_audio
  • candidate_3_images
  • candidate_3_audio
  • candidate_4_images
  • candidate_4_audio
modeReference (Omni) — up to 9 images, 3 videos, 3 audio
aspect_ratio16:9 (Widescreen)
megapixels0.98
multiple32
resize_methodcrop
duration_seconds10.0
chunk_duration_seconds10.0
ref_image_sizematch
hybrid_continuationfalse
seed0
seed_huntfalse
steps20
sampler_nameres_multistep
schedulersimple
shift_video12.00
shift_audio3.00
timeline_data{}
candidate_2false
candidate_3false
candidate_4false
CategoryMuse Collective

Inputs (25)

NameTypeDefaultDescription
modeCOMBOReference (Omni) — up to 9 images, 3 videos, 3 audio2 options: Reference (Omni) — up to 9 images, 3 videos, 3 audio, First/Last Frame — zero, one, or two frame images
modelMODEL
clipCLIP
vaeVAE
audio_vaeVAENeeded for final audio decode in both modes — H3 always builds a joint audio+video latent internally, even in First/Last Frame mode.
aspect_ratioCOMBO16:9 (Widescreen)8 options: 1:1 (Square), 2:3 (Portrait Photo), 3:2 (Photo), 3:4 (Portrait Standard), 4:3 (Standard), 9:16 (Portrait Widescreen), +2
megapixelsFLOAT0.980.1–4
multipleINT328–128
resize_methodCOMBOcropHow every character/background reference image and First/Last Frame image gets fit to the output resolution when its own aspect ratio doesn't match. 'crop' scales up and center-crops the excess (no distortion, may crop the edges of a person/scene). 'pad' scales down to fit entirely within the frame and adds black bars (nothing cropped, but the bars become visible reference content). 'stretch' resizes directly, distorting proportions.
duration_secondsFLOAT10.01–120Total length of the finished video. Automatically split into multiple H3 generation calls if longer than chunk_duration_seconds, stitched together.
chunk_duration_secondsFLOAT10.03–15Length of each individual H3 call. H3's own trained range tops out around 15s per call — longer totals get split into chunks this long (the final chunk absorbs whatever's left over, so it may be shorter). Reference mode: each continuation chunk is fed the previous chunk's own last few frames and last few seconds of audio as reference video/audio, plus an explicit instruction to continue seamlessly rather than cut. First/Last Frame mode: continuation falls back to the previous chunk's last frame only.
ref_image_sizeCOMBOmatch'match' scales references down to the generation's pixel area (faster). 'max' keeps up to a 2048px short edge for stronger identity fidelity, but reference tokens ride every sampling step so it's several times slower. Reference (Omni) mode only.
hybrid_continuationBOOLEANfalseReference (Omni) mode only, needs model_fl2va connected. Reference mode's own carry-over (ref_video/ref_audio) is a soft reference, not a hard lock — H3 can still cut to a new composition at a chunk boundary despite it. When this is on, continuation chunks (2nd onward) switch to a hard-locked first-frame anchor instead: the exact last frame of the previous chunk, via the separate First/Last-Frame checkpoint's real keyframe-lock mechanism. The first chunk always runs Reference (Omni) normally, so character/background images still establish identity — continuation chunks just don't get fresh reference-image reinforcement after that (the anchor frame itself already carries the correct likeness forward, since it's real output from the reference-anchored first chunk, not a blank start).
seedINT00–18446744073709550000
seed_huntBOOLEANfalseWhen on, runs 4 full passes total — identical settings, only the seed differs — and fills the candidate_1..4 outputs (candidate_1 is always the main seed; 2-4 use seed + N*1,000,003). Wire candidate_1..4_images/audio into MuseMinimaxRefine to pick one and refine it at higher resolution. Takes ~4x as long as a single run — set megapixels low here for cheap scouting, then refine at full resolution downstream.
stepsINT201–100
sampler_nameCOMBOres_multistep4 options: res_multistep, euler, euler_ancestral, dpmpp_2m
schedulerCOMBOsimple4 options: simple, normal, beta, sgm_uniform
shift_videoFLOAT12.000.01–100
shift_audioFLOAT3.000.01–100
timeline_dataSTRING{}
candidate_2BOOLEANfalseRuns one extra full pass (identical settings, seed + 1,000,003) and fills the candidate_2 output. Independent of Candidate 3/4 — turn on only the ones you want to pay for.
candidate_3BOOLEANfalseRuns one extra full pass (identical settings, seed + 2,000,006) and fills the candidate_3 output. Independent of Candidate 2/4.
candidate_4BOOLEANfalseRuns one extra full pass (identical settings, seed + 3,000,009) and fills the candidate_4 output. Independent of Candidate 2/3.
model_fl2vaoptMODELThe separate First/Last-Frame checkpoint (not the same weights as the main Reference/Omni model input) — load it via its own loader. Used whenever a First/Last-Frame-style generation actually happens: First/Last Frame mode itself, and Hybrid Continuation's chunk-to-chunk lock while in Reference mode. If left unconnected, First/Last Frame mode falls back to the main model input instead — which should normally hold the Reference/ref2va checkpoint, not this one, so results may be degraded.

Outputs (12)

NameTypeDescription
imagesIMAGE
audioAUDIO
compiled_promptSTRING
ref_images_usedIMAGE
candidate_1_imagesIMAGE
candidate_1_audioAUDIO
candidate_2_imagesIMAGE
candidate_2_audioAUDIO
candidate_3_imagesIMAGE
candidate_3_audioAUDIO
candidate_4_imagesIMAGE
candidate_4_audioAUDIO