foley.sources.stable_audio.adapter#

Stable Audio Open 1.0 generate adapter — local diffusers inference.

Turns a natural-language prompt into a GeneratedClip via diffusers.StableAudioPipeline running on-device (report 02 §Stable Audio Open). The model produces a 44.1 kHz stereo waveform (up to ~47 s); the adapter encodes it to container bytes, which the foley.sources.generate.generate() façade stores by-value (the generation flywheel) through the shared ingest_one() pipeline.

Load-bearing details verified against the diffusers docs + HF model card:

  • **The CFG kwarg is guidance_scale``** (NOT ``cfg_scale), and it MUST be passed explicitly or the pipeline silently uses its own default 7.0. foley maps its unified prompt_influence (0..1) onto it linearly (1..15, declared in the config’s coerce); None sends the model’s own default, and a distilled checkpoint (native_defaults["distilled"]) always gets its native default.

  • **Output is channel-first (channels, samples)``** — transposed to the time-first ``(frames, channels) foley/soundfile convention before encoding.

  • The sample rate is read off the pipeline (pipe.vae.sampling_rate = 44100), never hardcoded.

  • ``loop`` and ``output_format`` have no native pipeline param (loop is a future WEAVE crossfade; format is the audio-write layer) — they warn-and-drop; the adapter never fabricates a pipeline kwarg.

torch / diffusers are imported lazily inside the methods (the foley[stable-audio] extra), so import foley and source discovery stay dol-only. The local-model DI seam is pipeline= (the analog of the hosted adapters’ http=): inject a fake pipeline and no ML stack loads at all — the whole adapter is testable with no torch. The adapter performs NO storage/library access; the façade converges it on the shared ingest pipeline.

Module Attributes

Adapter

Registry convention (arioso): the loader imports adapter.Adapter.

Classes

Adapter

Registry convention (arioso): the loader imports adapter.Adapter.

StableAudioAdapter([config, pipeline])

Local Stable Audio Open 1.0 generate adapter (a GenerateAdapter).
foley.sources.stable_audio.adapter.Adapter#

Registry convention (arioso): the loader imports adapter.Adapter.

class foley.sources.stable_audio.adapter.StableAudioAdapter(config=None, *, pipeline=None)[source]#

Bases: object

Local Stable Audio Open 1.0 generate adapter (a GenerateAdapter).

Parameters:

  • config (Optional[dict]) – The SOURCE_CONFIG (defaults to the module’s). Passed positionally by the registry’s lazy loader (the arioso Adapter(config) convention).

  • pipeline – An optional pre-built pipeline (the dependency-injection seam — a test injects a fake callable exposing .vae.sampling_rate + .device; production omits it and the real diffusers.StableAudioPipeline lazy-loads on first generation).

generate(prompt, *, duration=None, prompt_influence=None, negative_prompt=None, steps=None, seed=None, loop=False, output_format='wav', **kw)[source]#

Generate a sound for prompt; return its bytes + provisional candidate.

Parameters:

  • prompt (str) – The natural-language sound description (required).

  • duration (Optional[float]) – Seconds (clamped to the model’s ~47.55 s max). None uses a sane 10 s default (NOT the 47.55 s maximum).

  • prompt_influence (Optional[float]) – 0..1 unified adherence, mapped linearly onto the native guidance_scale (1..15). None (default) sends the model’s own default (7.0 for Stable Audio Open 1.0); a distilled checkpoint ignores it with a note.

  • negative_prompt (Optional[str]) – Content to exclude (ignored by the model when guidance ≤ 1).

  • steps (Optional[int]) – Diffusion steps (native num_inference_steps); default 200.

  • seed (Optional[int]) – Reproducibility seed → a per-device torch.Generator; recorded in provenance. None → non-deterministic (no torch import).

  • loop (bool) – No native param (a future WEAVE crossfade) — warn-and-drop.

  • output_format (str) – No native param (the audio-write layer) — warn-and-drop; the clip is archived as FLAC by ingest regardless.

  • **kw – Extra/unknown affordances (warn-and-dropped).

Return type:

GeneratedClip

Returns:

A GeneratedClip (origin=generated, license_id='Stability-Community', is_ai_generated=True, generation_seed captured).