stable-audio-3-small-sfx-GGUF

Stable Audio 3 Small-SFX — GGUF (for sa3.cpp)

GGUF conversions of stabilityai/stable-audio-3-small-sfx for sa3.cpp — a portable C++/GGML port of Stable Audio 3, no PyTorch in the loop. Runs on CPU, CUDA, Vulkan, or Metal (Apple Silicon). The small-sfx model targets sound effects / foley (SAME-S autoencoder, 0.5B DiT). Validated against the PyTorch reference at cosine similarity ~1.0.

Files

Multi-file model. Grab the DiT + SAME at your chosen precision and the conditioner, plus the shared encoder + tokenizer from t5gemma-b-b-ul2-GGUF.

note: SAME-S needs an even --frames count (the packed sequence must divide the chunk size).

Encodings

sa3-generate --encoding resolves the DiT and the SAME with the same suffix, so download the pair.

q4_k_m and q5_k_m promote the attention V, feed-forward down and embedding tensors to Q6_K; q8_0 is uniform. Every tier passes sa3-quant-check with below-threshold=0 at cosine 0.990 against the F16 reference, for the DiT and the SAME alike.

Quantization buys footprint everywhere and speed only on some backends. CUDA and Vulkan gain roughly 33% end to end. Metal is flat — Q8_0 is 1.7% faster and Q4_K_M 2.1% slower than F16, because the load-time saving and the added per-step dequant cancel out. On a Mac, pick a quant for the memory, not for the speed.

Usage

For use with sa3.cpp:

python tools/download_models.py --variant small-sfx --encoding f16

# --model resolves the gguf set in ./models by name
sa3-generate --model small-sfx --prompt "a dog barking in a large empty hall" --out sfx.wav

For a quantized set, pass the encoding to both — the downloader and the generator use the same names:

python tools/download_models.py --variant small-sfx --encoding q4_k_m
sa3-generate --model small-sfx --encoding q4_k_m --prompt "a dog barking in a large hall" --out sfx.wav

Performance

Roughly 1.7s for a 12s clip at f16 on an 8GB laptop GPU (RTX 5070) — about 2× faster than the medium model. The sliding-window decoder keeps long generations linear. Full numbers + levers: docs/BENCHMARKS.md.

License

These are format conversions of stabilityai/stable-audio-3-small-sfx, whose weights Stability AI releases under the Stability AI Community License: free for organizations under $1M annual revenue, with commercial use, fine-tuning, and derivative works permitted within that threshold (above it, contact Stability AI for an Enterprise License). Outputs are yours. That license carries over to these converted weights.

The upstream stable-audio-3 source code is released separately under MIT. Pair these with the shared T5Gemma text encoder, which is Google's under the Gemma Terms of Use.

Relationship to the original

Format conversions (weights → GGUF) for inference in sa3.cpp — no retraining. See sa3.cpp/docs/DISTRIBUTION.md.