Nodes/Muse-MiniMax-H3-Refine/Muse Minimax Refine
ComfyUI Node

Muse Minimax Refine

A ComfyUI node in Muse Collective with 27 inputs and 2 outputs.

By muse-collective-26·Created 19 days ago·Updated 19 days ago· 2
Muse Minimax Refine
  • model
  • clip
  • vae
  • audio_vae
  • candidate_1_images
  • candidate_1_audio
  • candidate_2_images
  • candidate_2_audio
  • candidate_3_images
  • candidate_3_audio
  • candidate_4_images
  • candidate_4_audio
  • ref_images
  • images
  • audio
prompt
candidate0
aspect_ratio16:9 (Widescreen)
megapixels1.96
multiple32
resize_methodcrop
ref_image_sizematch
denoise0.40
seed0
steps20
sampler_nameres_multistep
schedulersimple
shift_video12.00
shift_audio3.00
CategoryMuse Collective

Inputs (27)

NameTypeDefaultDescription
modelMODEL
clipCLIP
vaeVAE
audio_vaeVAE
promptSTRINGWire this from your director-style node's compiled_prompt output — the refine pass reuses the exact prompt the candidate was generated from.
candidateINT00–4Which of the four candidate slots to refine. Set by the button selector in the node's UI. Defaults to 0 (none picked yet) — the node deliberately refuses to run at 0, rather than silently falling back to candidate 1, so a graph can't get queued before you've actually chosen one.
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
megapixelsFLOAT1.960.1–4Target resolution for the refined output — normally higher than whatever resolution the candidates were scouted at.
multipleINT328–128
resize_methodCOMBOcropHow the selected candidate's frames get fit to the new target resolution when its aspect ratio doesn't match.
ref_image_sizeCOMBOmatchOnly used when ref_images is connected. 'match' scales references down to the output'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.
denoiseFLOAT0.400.05–1How much of the candidate's own content survives vs. gets redrawn. Low values (~0.2-0.3) stay close to the candidate and mostly add resolution/detail. High values (~0.6+) let the model redraw more, which can drift from the candidate's content — 1.0 regenerates from scratch and ignores it entirely.
seedINT00–18446744073709550000
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
candidate_1_imagesoptIMAGE
candidate_1_audiooptAUDIO
candidate_2_imagesoptIMAGE
candidate_2_audiooptAUDIO
candidate_3_imagesoptIMAGE
candidate_3_audiooptAUDIO
candidate_4_imagesoptIMAGE
candidate_4_audiooptAUDIO
ref_imagesoptIMAGEThe same character/product reference photos that anchored identity and fine detail (exact prop shape, skin, likeness) in the original candidate — e.g. wired from MuseMinimaxDirector's ref_images_used output. Without this, the refine pass only has the text prompt to go on, and any detail the prompt doesn't spell out explicitly is free to drift during the partial-denoise redraw — this is what locks it back down.

Outputs (2)

NameTypeDescription
imagesIMAGE
audioAUDIO