---
license: other
license_name: stability-ai-community-license
license_link: https://huggingface.co/stabilityai/stable-audio-3-small-sfx/blob/main/LICENSE.md
tags:
  - audio
  - text-to-audio
  - sound-effects
  - stable-audio
  - onnx
  - q4
  - matmulnbits
  - onnxruntime-web
  - webgpu
library_name: onnxruntime
base_model: stabilityai/stable-audio-3-small-sfx
---

# stable-audio-3-small-sfx — ONNX (browser / WebGPU)

The **SFX-specific** ONNX graphs for [`stabilityai/stable-audio-3-small-sfx`](https://huggingface.co/stabilityai/stable-audio-3-small-sfx),
for text-to-sound-effect generation in the browser via [`onnxruntime-web`](https://www.npmjs.com/package/onnxruntime-web).

This model shares its **text encoder** (T5Gemma) and **autoencoder/decoder** with the music
model — those weights are bit-identical — so only the SFX-specific parts live here. Reuse the
text encoder and decoder from
[`lsb/stable-audio-3-small-music-onnx`](https://huggingface.co/lsb/stable-audio-3-small-music-onnx).

## Files

```
dit_q4.onnx + dit_q4.data     SFX diffusion transformer, int4 MatMulNBits (~320 MB)
number_conditioner.onnx       SFX duration embedder
padding_embedding.json        SFX prompt padding vector (768 floats, see below)
```

Linear/MatMul weights are int4 `MatMulNBits` (block_size 32); everything else (Conv, norms,
the embedding `Gather`) stays fp32 — so every op has a WebGPU kernel and the graph runs on the
`webgpu` execution provider (with `wasm` fallback).

## Pipeline (identical to the music model)

```
tokens → text_encoder(lsb) → [override pad rows with padding_embedding]
seconds → number_conditioner → duration token
cross = [text(256); duration(1)] (1,257,768);  global = duration (1,768)
pingpong sampler, 8 steps, LogSNRShift(rate=0, anchor_logsnr=-6.2, logsnr_end=2.0), CFG off
latents → decoder(lsb) → stereo 44.1 kHz
```

### DiT I/O

```
x               float32 (1, 256, t_lat)
t               float32 (1,)
cross_attn_cond float32 (1, 257, 768)
global_embed    float32 (1, 768)
local_add_cond  float32 (1, 257, t_lat)   zeros for plain text-to-audio
padding_mask    bool    (1, t_lat)        all-true
→ out           float32 (1, 256, t_lat)
```

`t_lat = ceil((seconds + 6) · 44100 / 8192) · 2`.

### padding_embedding

The prompt conditioner replaces padded token positions (where `attention_mask == 0`) with a
learned vector. That vector differs between the SFX and music models, so after running the
(shared) text encoder, overwrite the padded rows of `last_hidden_state` with this 768-vector.

## License

Inherits the **Stability AI Community License**; the T5Gemma text encoder additionally falls
under Google's Gemma Terms of Use.