Prompt Stable Audio for sound effects

Last updated: October 12, 2026

Stable Audio 3 Small SFX turns a description of a sound into a one-shot or an ambience at the length you set. Pick it for game foley, impacts, pickups and UI sounds when you can name what makes the sound, what it does, and how it was recorded.

Let the agent write it

You don't have to write any of this yourself. The Layer agent knows how to prompt each model it runs, from the same guidance this page is built on. Name the model and the sound, for example: Make a heavy castle gate slamming shut with Stable Audio 3 Small SFX, dry, about three seconds. Give me a few takes. The agent writes the prompt, shows you its plan, and runs it when you approve. Use the rest of this guide when you want to write or adjust the prompt yourself, and see Write a brief for the agent for how to brief it.

When to use it

Use Stable Audio 3 Small SFX for single game sounds: doors and gates, weapon hits, footsteps, coin and item pickups, menu sounds, and short ambiences. It is made for sound effects only, so it is not the model for music or for voices.

For a soundtrack or a menu loop with a melody, see Create music. For a sequence of events, make each sound on its own and put them together in an audio editor.

Versions on Layer

Stable Audio 3 Small SFX

Made for sound effects and ambience. Each run returns one MP3. Start here for any game sound.

Stable Audio 2.5

Stability AI's general model, built and documented mainly for music, with room for long tracks. It also makes a one-shot from the same kind of prompt, and returns a WAV file. In our test it made one gate slam from the prompt below, as Small SFX did.

How to write the prompt

The source, the action and how it decays, then how it was recorded:

A heavy iron portcullis slamming shut onto stone. A pronounced low-mid body with a short metallic rattle and a fast decay. Close-mic'd in a dead room, no reverb.

Say exactly what makes the sound, then how it is triggered and how long it rings. End with the production: where the mic is, what the room is like, and any processing. The model learned from sound libraries and their descriptions, so it reads the words a sound designer would put in a file's notes. That prompt is the one in the clip above.

What works

A duration that fits the sound

Set Duration before you run. It starts at 30 seconds on Stable Audio 3 Small SFX and 190 seconds on Stable Audio 2.5, and the model fills the length it gets. Left at 30 seconds, our gate prompt came back as four separate slams spread across the clip. At 3 seconds it came back as one slam, with a little over a second of silence after it to trim.

A seed to keep a take

The same Seed, prompt and settings give the same file: we ran one request twice and got identical audio. When a take is close, keep its seed and change one word or one setting at a time. Leave the seed empty to get a new take on every run.

Source, action, production

Name the object, then the action, then the recording. Stability AI's own examples read like this: "A blunt, powerful thud made by slamming a wooden desk drawer shut. It has a pronounced low-mid body, making it feel heavy, and is given a touch of analog distortion for aggressive character."

Recording words

Mic placement, room and processing all steer the sound: "close-mic'd", "recorded in a dead room", "captured with a vintage ribbon mic", "a dark wooden reverb", "a touch of analog distortion". Many of these terms are in the descriptions the model was trained on.

Say how it decays

The tail is part of the action. "Zero ring and a very fast decay" and "a short metallic rattle" tell the model how long the sound should last after the hit.

What to avoid

Raising Guidance scale

Stability AI suggests a higher guidance for stronger prompt adherence, but only on its base models; Small SFX is not one. On Layer, our gate prompt at Guidance scale 7 came back as three seconds of sound at full level with no silence and no decay, where the default of 1 gave one clean slam. Leave it at 1.

Leaning on the negative prompt

Stability AI's documentation says its sound-effects model ignores a negative prompt, as it does guidance. Say what you want in the prompt instead: "dry, close-mic'd, dead room" rather than a negative prompt of "reverb".

Prompts in other languages, or speech

The model was trained on English descriptions and is not made for voices. Write the prompt in English, and use a speech model for dialogue.

Settings that matter

Duration sets the clip length; set it close to the sound for every one-shot. Seed repeats a take exactly. Guidance scale and Inference steps are best left at 1 and 8: more steps do not add quality on Small SFX, and fewer trade quality for speed.

Prompts to start from

Heavy gate slam

A heavy iron portcullis slamming shut onto stone. A pronounced low-mid body with a short metallic rattle and a fast decay. Close-mic'd in a dead room, no reverb.

Run with Duration at 3 seconds.

Coin pickup

A single magic coin pickup for a fantasy game: a bright glassy chime with a soft sparkle shimmer on top. Short and clean with a fast decay, recorded dry in a dead room.

Run with Duration at 2 seconds; the chime is over in about one.

Use it in Layer

  1. In a session, open Create sound effects from the side panel.

  2. Click the model name at the top of the panel and choose Stable Audio 3 Small SFX.

  3. In the prompt box, describe the source, the action and its decay, and how it was recorded.

  4. Set Duration to a little longer than the sound.

  5. Click Create sound effects. The cost shows on the button before you click.

  6. When the sound lands on the board, click View to play it. Run again for more takes, and note the seed of the one you keep.

Learn more

  • Prompt Guide (Stable Audio 3) (Stability AI): the source, action and production structure for sound effects, with worked prompts, and why a short duration fits.
  • Inference (Stable Audio 3) (Stability AI): what steps, seed, guidance and negative prompt do, and which of them the post-trained models ignore.
  • Model overview (Stable Audio 3) (Stability AI): what each Stable Audio 3 model is for, its length limit, and its limits on speech and non-English prompts.
  • Stable Audio 2.5 prompt guide (Stability AI): how Stable Audio 2.5 reads a prompt, with production and era vocabulary it understands.

Related