Nodes/M3 SongPlanner/M3 Song Planner (Local VLM)
ComfyUI Node

M3 Song Planner (Local VLM)

Turns a creative idea into a production-ready MiniMax Music 3 caption + lyrics pair using a LOCAL VLM (Qwen3-VL / Gemma-3-Vision). Two-stage pipeline: caption generation → lyrics generation. No API keys needed.

By benjiyaya·Created 19 days ago·Updated 19 days ago· 9
M3 Song Planner (Local VLM)
  • clip
  • caption
  • lyrics
  • debug
text_encoder
idea
genre_hint
vocal_configauto (decide from idea)
languageEnglish
duration_seconds120
seed0
temperature0.80
top_p0.95
top_k64
max_tokens2048
keep_model_loadedtrue
CategoryM3/SongPlanner

Inputs (13)

NameTypeDefaultDescription
text_encoderCOMBOVLM checkpoint from models/text_encoders/ (e.g. Qwen3-VL or Gemma-3-Vision instruct repack).
ideaSTRING
genre_hintSTRING
vocal_configCOMBOauto (decide from idea)5 options: female vocals, male vocals, duet, instrumental, auto (decide from idea)
languageCOMBOEnglish5 options: English, Chinese (Mandarin), Korean, Japanese, auto
duration_secondsFLOAT12030–300Target song length. Affects structure and token budget.
seedINT00–4294967295
temperatureFLOAT0.800–2
top_pFLOAT0.950–1
top_kINT640–500
max_tokensINT2048512–8192Max tokens per generation (caption and lyrics each).
keep_model_loadedBOOLEANtrueCache the text-encoder CLIP between runs.
clipoptCLIPConnect CLIP from CLIPLoader. Overrides the dropdown.

Outputs (3)

NameTypeDescription
captionSTRINGWire to MiniMaxMusic3TextEncode 'caption' input.
lyricsSTRINGWire to MiniMaxMusic3TextEncode 'lyrics' input.
debugSTRINGJSON with timings, clip source, model family, warnings.