Stable Audio 3

Stable Audio 3

A state-of-the-art open platform for fast, high-quality generated audio and music.

Technical Report · 🤗 Models · 🤗 Extra Models · Discord · Demo · Blog Post

Stable Audio 3 is the next generation of Stable Audio: a focused, streamlined platform for inference and fine-tuning, built on lessons from stable-audio-tools. If you're doing foundational research or working with previous Stable Audio models, that repo is still the place to go.

Models

| Model | Model ID | Autoencoder | Hardware | Params | Max length | Use case |
| Stable Audio 3 Small-Music | small-music | SAME-Small | CPU | 433M | 120s | Lightweight music-only inference, no GPU required |
| Stable Audio 3 Small-SFX | small-sfx | SAME-Small | CPU | 433M | 120s | Lightweight sound effects-only inference, no GPU required |
| Stable Audio 3 Medium | medium | SAME-Large | GPU (CUDA) | 1.4B | 380s | High Quality, Fast Inference |
| Stable Audio 3 Large | — | SAME-Large | API only | 2.7B | 380s | Highest quality, API only. Not supported by this repo, see the API docs |

Base (un-post-trained) checkpoints, the SAME autoencoders, and optimized variants are available in the Extra Models collection.

Performance

| Model | Duration | H200 | H200 + TensorRT | Mac CPU* | Mac CoreML | Peak VRAM† |
| small | 5s | 0.41s | 0.017s | 0.70s | 0.23s | 1.69 GB |
| small | 30s | 0.46s | 0.022s | 1.72s | 0.63s | 1.89 GB |
| small | 120s | 0.45s | 0.044s | 5.92s | 3.09s | 2.40 GB |
| medium | 5s | 0.60s | 0.02s | – | – | 5.07 GB |
| medium | 30s | 0.65s | 0.05s | – | – | 5.49 GB |
| medium | 120s | 0.78s | 0.13s | – | – | 6.49 GB |
| medium | 380s | 1.31s | 0.43s | – | – | 6.52 GB |

\* CPU-only via CoreML (Diffusion Transformer) + TFLite (SAME-S decoder)
† Peak allocated VRAM on H200, unchunked decode. Chunked decoding reduces this — e.g. medium at 120s drops from 6.49 GB to ~5.14 GB.

Features
- ⚡ Fast, state-of-the-art generation - Generate minutes of audio in milliseconds
- 🎛️ Three inference modes — text-to-audio, audio-to-audio editing, and inpainting/continuation
- ↔️ Variable-length generation — handles generation of a variety of sequences without wasting inference time and VRAM on unused latents
- 🎯 Personalization through LoRA fine-tuning — adapt any model to a target style; stackable, adjustable at runtime
- 💻 Broad hardware support — CPU (Small), CUDA/TensorRT (Medium), Apple Silicon via CoreML, Others coming soon
- 🎵 SAME autoencoder — new Semantic-Acoustic Music Encoder; stereo, 44.1 kHz, 256-dimensional latents optimized for both generative tractability and high-quality reconstruction

Installation

Stable Audio 3 uses uv for fast, lightweight installs. Install only what you need.

bash
Base install (Python API only)
uv sync

With Gradio UI
uv sync --extra ui

With LoRA training support
uv sync --extra lora

Everything
uv sync --extra ui --extra lora

CUDA Version

By default, uv sync installs PyTorch built against CUDA 12.6. If you need a different CUDA version, install torch and torchaudio manually first (pinning the same version as pyproject.toml), then sync without reinstalling them, for example:

bash
uv pip install torch==2.7.1 torchaudio==2.7.1 --index-url https:/

……（后略，见原文）