Stable Audio 3 in ComfyUI: setup and audio workflows

This tutorial uses Stable Audio 3 Medium in ComfyUI to generate instrumental music and sound effects on local hardware. Setup requires an updated ComfyUI installation, an approximately 8 GB audio model in the checkpoints folder, and a T5 Gemma model in text_encoders. The Load CLIP node uses the stable audio option.

The speaker describes a duration range of 1 to 380 seconds and says the model does not generate lyrics. Most tests use 30-second clips. On the speaker's 24 GB VRAM card, a repeat generation takes under two seconds after model loading. That result is specific to the demonstrated setup. Longer clips can spend much more time in VAE decoding; the speaker suggests trying the tiled audio decoder when VRAM is limited.

Further workflows use Gemma 4 to expand short descriptions into music or sound-effect prompts. The extra model is about 8 GB and increases VRAM needs when running alongside the audio model. An experimental image workflow asks Gemma 4 to write a music prompt based on a picture, then passes that text to Stable Audio 3.

The final section covers Pixaroma Nodes updates: color copying, four favorite colors, image previews and filtering, plus manual padding controls. The speaker also advises clearing the browser cache if updated node controls fail to appear.