CoolFace
Modelpublic

PulpCut/Stable-Audio-3-Small-SFX-safetensors

sourceHugging Faceotherupdated 1mo agoView on Hugging Face
0likes
Model Card

Stable Audio 3 Small SFX — safetensors

Powered by Stability AI.

This is Stability AI's Stable Audio 3 Small SFX repackaged into safetensors for H3ddle, an open-source native macOS app that generates audio locally on Apple Silicon.

Nothing about the model was changed. Tensor names, shapes and values are identical to the source; only the container differs. No retraining, merging, pruning or quantization was performed.

Why this exists

Stability publishes these weights twice: as fp32 in the gated stable-audio-3-small-sfx repository, and as f16 NumPy archives in the ungated `stable-audio-3-optimized` repository. This repackages the second, because it needs no Hugging Face account and is half the size. The encoder half of the autoencoder is omitted: text-to-audio never runs it, and leaving it out saves 215 MB.

Safetensors is what the consuming engine already reads, so this avoids teaching an installer to parse .npz.

Contents

filewhat it is
dit.safetensorsthe 20-block diffusion transformer, f16
text_encoder.safetensorsT5Gemma encoder, f16
decoder.safetensorsSAME-S audio decoder, f32 as released
tokenizer.jsonHugging Face fast-tokenizer form
tokenizer.modelthe original SentencePiece model

The text encoder's architecture travels in the safetensors metadata under config, where it shipped as a JSON blob in the source archive.

Measured H3ddle performance versus the regular CPU path

The safetensors conversion does not itself make inference faster: all 690 included tensors round-trip bit-identically.

comparisonregular / baselineoptimized pathresulting gain
same 20-block transformer forward on M1 Pro1.00x CPU14.0x Metal14x stage speed
one-time text encoderabout 0.1 s CPUabout 0.1 s CPUdeliberately unchanged

The transformer runs eight times per default generation, while the one-time text encoder remains on CPU because its cost is already small. Complete-generation gain varies with requested audio length and decoder share.

Licence

The model is licensed under the Stability AI Community License, Copyright © Stability AI Ltd. All Rights Reserved. It is free for research and non-commercial use, and for commercial use below USD $1M in annual revenue after registering with Stability AI. See NOTICE for the required attribution and a statement of the modifications made here.

The text encoder is T5Gemma, provided under and subject to the Gemma Terms of Use; see LICENSE_GEMMA.md.

Outputs may not be used to train competing foundational generative models.