Stable Audio 3 in ComfyUI: low-VRAM music tutorial

This tutorial builds a ComfyUI workflow for Stable Audio 3 music and sound effects on home hardware. The presenter compares Medium, Small and Medium Base through audio examples. Medium uses eight sampling steps at CFG 1; Small is described as a 2 GB model with lower output quality. Medium Base uses 50 steps and a higher CFG. The presenter finds its results repetitive and suggests the undistilled model for fine-tuning experiments.

The workflow includes duration controls and an optional model shift node. The presenter explains that negative prompts do not work at CFG 1, then compares Medium output with and without a shift value of one. For automated prompting, Gemma 4 describes an audio clip, and that description becomes the music prompt.

Audio-input examples use VAE encoding, including a clip made with ACE-Step and a recording of the presenter singing. The resulting music differs from both the input and generations made without it. The presenter notes that this model cannot produce vocals. Sound-effect tests include creaky doors and underwater fireworks.

The final section splits an eight-step generation between two samplers at step four to adjust the result. The presenter says their ComfyUI workflow lacks the audio inpainting available in the Gradio interface.