MiniMax Music 3 in ComfyUI: local prompt setup

This tutorial combines MiniMax Music 3 with Pixaroma prompt nodes in ComfyUI. The speaker explains how to expand short ideas into image prompts, then generate separate music captions and lyrics with a local LLM. Examples cover Krea 2, prompts derived from images, and Z Image Turbo.

The prompt presets pair instructions with specific models. The speaker warns that another model version may respond differently, even when its name is similar. Settings include temperature, fixed or random seeds, and custom presets. A free VRAM option unloads the prompt model after generation, with a speed penalty. The Krea 2 example reuses the workflow's model rather than loading it twice.

For music, the workflow connects caption and lyrics outputs to MiniMax Music 3 and shares a duration control. The prompt node accepts custom durations up to 300 seconds, but the speaker stresses that duration is a maximum, not a promised song length. Tests produce early endings and cut-off lyrics; changing the seed can alter the length even with identical text. Fixed lyrics allow manual edits, while dynamic tiled audio decoding is suggested when longer songs get stuck.

The speaker uses a 24 GB VRAM card, without establishing it as a minimum requirement. The final sections demonstrate ChatGPT-generated music prompts, discuss Gemini and Claude instructions, and show RunningHub as a cloud option.