Skip to content

Text to Sound

Audio and song composition built on MiniMax Music 3 GGUF quants.

Setup

Music generation replaces the pixel or frame VAE with an audio decode stage. The loading pattern is the same: load the pruned text encoder and the minimax_music3_dit backbone with the GGUF loader nodes from comfyui-gguf-reboot.

Use the GGUF loader nodes for the best performance.

Models and formats

ModelRepositoryRecommended formats
MiniMax Music 3molbal/Minimax-Music3-GGUFQ8_CR, Q8_0, Q4_0

Loading instructions

  1. Download the MiniMax Music 3 GGUF files from Hugging Face.
  2. Load the pruned text encoder and minimax_music3_dit backbone with the GGUF loader nodes.
  3. Set the audio length, up to 5 minutes, and the lyrics or description conditioning.
  4. Configure the audio sampler and decode to a waveform.
  5. Export the audio file.

Quality

Use Q8_CR for music. ConvRot keeps sensitive audio layers precise within INT8 memory bounds. See MiniMax Music 3.

Showcase

Audio examples are on molbal's Civitai profile.

GGUF quantization and open-source tools.