README.md

---
license: other
license_name: stability-ai-community-license
license_link: https://stability.ai/license
library_name: stable-audio-3
tags:
  - audio
  - audio-generation
  - text-to-audio
  - audio-to-audio
  - inpainting
  - stable-audio-3
  - stability-ai
  - safetensors
pipeline_tag: text-to-audio
---

# Stable Audio 3 — bundled mirror

Self-contained inference bundle for the [MAESTRO](https://github.com/AEmotionStudio/MAESTRO) desktop app.
One-to-one mirror of Stability AI's [Stable Audio 3 collection](https://huggingface.co/collections/stabilityai/stable-audio-3) and the [extras collection](https://huggingface.co/collections/stabilityai/stable-audio-3-extra) (base checkpoints + standalone autoencoders), bundled into a single browseable HF repo so the MAESTRO panel can pick the variant a user wants without juggling eight separate downloads.

## License — Stability AI Community License

All weights in this repository are released by Stability AI under the **[Stability AI Community License](https://stability.ai/license)**:

> Free for organizations with **under $1M annual revenue**. Commercial use of the models and outputs is permitted within that threshold; redistribution, fine-tuning, and derivative works are explicitly allowed. **Outputs are yours.** Above the revenue threshold, contact Stability AI for an Enterprise License.

The upstream [`stable-audio-3` source code](https://github.com/Stability-AI/stable-audio-3) is released separately under **MIT**.

### Gated subdirs

Three subdirs mirror upstream repos that are **gated** on huggingface.co — you must accept Stability AI's terms (and the Gemma terms-of-use, since the text encoder is T5-Gemma) before this mirror's gating allows access:

- `small-music/` (mirror of [`stabilityai/stable-audio-3-small-music`](https://huggingface.co/stabilityai/stable-audio-3-small-music))
- `small-sfx/`   (mirror of [`stabilityai/stable-audio-3-small-sfx`](https://huggingface.co/stabilityai/stable-audio-3-small-sfx))
- `medium/`     (mirror of [`stabilityai/stable-audio-3-medium`](https://huggingface.co/stabilityai/stable-audio-3-medium))

The base checkpoints and SAME autoencoders are open.

## Contents

| Subdir | Role | Params | Max duration | Upstream |
|---|---|---|---|---|
| `small-music/` | Post-trained text → audio (music) | 433 M | 120 s | `stabilityai/stable-audio-3-small-music` *(gated)* |
| `small-sfx/` | Post-trained text → audio (SFX) | 433 M | 120 s | `stabilityai/stable-audio-3-small-sfx` *(gated)* |
| `medium/` | Post-trained text → audio (music + SFX) | 1.4 B | 380 s | `stabilityai/stable-audio-3-medium` *(gated)* |
| `small-music-base/` | Base ckpt for LoRA fine-tuning | 433 M | 120 s | `stabilityai/stable-audio-3-small-music-base` |
| `small-sfx-base/` | Base ckpt for LoRA fine-tuning | 433 M | 120 s | `stabilityai/stable-audio-3-small-sfx-base` |
| `medium-base/` | Base ckpt for LoRA fine-tuning | 1.4 B | 380 s | `stabilityai/stable-audio-3-medium-base` |
| `same-s/` | SAME-Small standalone autoencoder | ~50 M | — | `stabilityai/SAME-S` |
| `same-l/` | SAME-Large standalone autoencoder | ~200 M | — | `stabilityai/SAME-L` |

Every subdir contains `model.safetensors` + `model_config.json` (plus the post-trained / base variants include the bundled T5-Gemma text encoder + SAME pretransform; SAME repos are AE-only).

## Capabilities

All six generative variants share a single inference surface in MAESTRO with four modes:

- **Text → Audio** — prompt-only generation, stereo 44.1 kHz
- **Audio → Audio** — style transfer / restyling with an adjustable `init_noise_level`
- **Inpaint** — multi-region regeneration of a source clip; non-region time is preserved verbatim
- **Continue** — extend an existing clip past its end

Generation knobs exposed: prompt, negative prompt, duration, steps, CFG scale, APG scale, seed, batch size, sampler type (`dpmpp-3m-sde` / `dpmpp-2m` / `euler` / `heun`), distribution shift (`logSNR` / `flux` / `identity`), precision (fp16 / fp32), chunked decode, and a user-loadable stackable LoRA stack.

> **Medium variants** require **[Flash Attention 2](https://github.com/Dao-AILab/flash-attention)** for the SAME-Large decoder path. Without `flash-attn` installed, Medium generation degrades to static-glitch output. Small variants do not require it.

## Format

- **All weights are `safetensors`.** No `.pt` / `.ckpt` / `.bin` in this mirror.
- Mirror is **bf16** — re-saved via `safetensors.torch.save_model` (preserves shared RotaryEmbedding buffers that bare `save_file` would corrupt). Bytewise this halves disk size vs the fp32 upstream. The MAESTRO runner upcasts to fp32 transiently during `load_state_dict` then casts to fp16 (`model_half=True`) for inference — runtime VRAM is unchanged from the fp32 mirror, but disk + I/O + initial safetensors-read CPU spike are all halved.
- Approximate disk sizes per subdir: small variants ~1.14 GB each, medium variants ~4.61 GB each, SAME-S ~0.22 GB, SAME-L ~1.70 GB. Total mirror footprint ≈ 15.7 GB.

## Usage

### Inside MAESTRO

The MAESTRO desktop app's `AI > Create > Stable Audio 3` panel handles the download + variant selection. The bundled runner at `backend/ai/models/stable_audio_3.py` reads the per-variant subdir name from the manifest and feeds it into the vendored `stable_audio_3` package at `backend/ai/stable_audio_3_vendor/`.

### Standalone

The repo can also be consumed directly by Stability AI's upstream [`stable-audio-3` package](https://github.com/Stability-AI/stable-audio-3):

```python
from stable_audio_3.loading_utils import load_diffusion_cond
from stable_audio_3.model import StableAudioModel
import json
from huggingface_hub import snapshot_download

# Pull one variant (e.g. small-sfx)
local = snapshot_download(
    repo_id="AEmotionStudio/stable-audio-3-mirrors",
    allow_patterns=["small-sfx/**"],
)

with open(f"{local}/small-sfx/model_config.json") as f:
    cfg = json.load(f)

inner = load_diffusion_cond(cfg, f"{local}/small-sfx/model.safetensors",
                            device="cuda", model_half=True)
inner.use_lora = False
inner.lora_names = []
model = StableAudioModel(inner, cfg, "cuda", model_half=True)

audio = model.generate(
    prompt="heavy rain on a tin roof with distant thunder",
    duration=10,
    steps=8,
    cfg_scale=1.0,
)
```

## Attribution

- **Models:** Stability AI — *Stable Audio 3* ([blog](https://stability.ai/news/stable-audio-3-open), upstream code: [`Stability-AI/stable-audio-3`](https://github.com/Stability-AI/stable-audio-3)).
- **Text encoder:** Google T5-Gemma (bundled in each generative subdir).
- **Autoencoder:** Stability AI SAME — *Semantic-Acoustic Music Encoder*.

This mirror exists to bundle the family + extras into a single browseable HF repo for the MAESTRO desktop app. It does not modify the weights; report quality or licensing issues to the upstream repos.