Apppublic

sleeper371/stable-audio-3-medium-api

sourceHugging Faceotherupdated 8d agoView on Hugging Face

0likes

App README

Stable Audio 3 Medium — Music API

A thin, stable API wrapper around `stabilityai/stable-audio-3`, pinned to the Medium — general audio variant (`stabilityai/stable-audio-3-medium`).

This Space uses no GPU of its own — it forwards to the upstream ZeroGPU Space and exposes a single keyword-addressable endpoint, /generate, so callers don't have to track upstream's positional, UI-derived signature.

Endpoint

POST /generate → (wav_filepath, status_string)

param	type	default	notes
`prompt`	str	—	required
`duration`	float	60	seconds, 1–380
`steps`	float	8	diffusion steps
`cfg_scale`	float	1.0
`sampler_type`	str	`pingpong`	`pingpong`, `euler`, `rk4`, `dpmpp`
`seed`	int	0	0 = random
`negative_prompt`	str	`""`	routes via upstream `/infer_advanced` when set
`hf_token`	str	—	required; your own HF token (a read token suffices)

Usage

python

from gradio_client import Client

client = Client("sleeper371/stable-audio-3-medium-api")
wav_path, status = client.predict(
    prompt="Warm analog house groove, deep sub bass, shuffled hats 122 BPM",
    duration=30,
    hf_token="hf_...",      # your token — your quota
    api_name="/generate",
)
print(status, wav_path)

Configuration

This Space stores no credentials. Each caller passes their own hf_token, which is forwarded to upstream so ZeroGPU quota is charged to that caller — never to the Space owner. Tokens are never logged or persisted.

variable	purpose
`UPSTREAM_SPACE`	Override the upstream Space id. Defaults to `stabilityai/stable-audio-3`.
`JOB_TIMEOUT`	Wall-clock budget per upstream job, including queue wait. Defaults to `900` seconds.
`HTTPX_TIMEOUT`	HTTP timeout, which bounds the download of the result WAV (~67 MB at the 380s maximum). Defaults to `600` seconds.