Video - https://youtu.be/X-7COOaRuQc

Overview

Stability AI? Didn't they do... something a while back? Well, they've done another thing - Stable Audio 3.

What does it do? Music. No lyrics - just the music.

Small and Medium models are available for use in ComfyUI, with their GitHub repo also featuring a Gradio interface instead, along with some extra features

Github: https://github.com/Stability-AI/stable-audio-3

Even the medium model is small, but the small model is under 3GB and is so fast you can run it using just your CPU!

The music quality is great, and if you had trouble with AceStep creating hardcore techno beats, then give Stable Audio 3 a try!

It isn't only music though, as it can create a range of sound effects too. No more hunting in sample libraries to find that door creak or lightning crack you're looking for - you can generate samples yourself!

Sound effects are generally just 1 or 2 seconds long, with music up to around 6 minutes long.

I suggest using the Medium (non-base) model for everything, but go ahead and try any model you like!

Workflow Information

The attached workflow works with all checkpoints - just remember to update the KSampler settings if using "base".

Model Downloads

There are a variety of models you can download. The base models need higher CFG and steps, so I'd instead start with the distilled Medium model as it give you something closer to what you'd likely want without having to mess around.

The workflow works with of of the models - just remember to change the CFG/Steps if using a base model!

Custom Nodes

The custom node requirements are minimal - just KJ and RG:

Loader Group

The models are loaded using the checkpoint loader, and as the shift node seems to have an interesting impact, I've got an option for that in there too:

Prompting Group

You'll mostly just want to use the positive prompt, unless using a base model:

Prompts can be short and sweet, but don't try to include lyrics unless you want an ... interesting experience. BPM can be included in the prompt.

If you enable the Load Audio group, then you can pick some input audio:

Interestingly enough using input audio - even with a full denoise - will change the generation quite considerably. Their repo uses a denoise value of 0.9 and your mileage may vary when using an audio reference in ComfyUI.

If you don't want to type prompts yourself, but instead want an AI to look at some audio and describe it for you, then the AI Prompting group can be enabled:

This uses Gemma4 to describe the audio and produce a prompt suitable for Stable Audio 3.

Sampling Groups

The main sampler group is where you can set the audio length (if not loading audio) and change the Ksampler settings. A note is included with the extra info:

For the most part only the LCM sampler makes anything decent, but do play around and let me know if you find any other decent combos!

For extra fun you can also trying the double sampler group:

This way you can subtly (or totally) change the result by using a second sampler.

Enjoy! :)