File size: 2,405 Bytes
b2e4883 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | #!/usr/bin/env python3
"""
SA3 audio-to-audio spike, mirroring the official HF Space's default settings.
The Space (stabilityai/stable-audio-3) uses:
steps=8, cfg_scale=1.0, sampler=pingpong, sigma_max=1.0,
apg_scale=1.0, duration_padding_sec=6.0, seed=-1,
init_noise_level=0.9 (Advanced -> Init audio)
High cfg_scale (e.g. 4.0) over-processes noisy input into garbage; the Space
defaults to cfg_scale=1.0 which is far gentler.
"""
import argparse
import torchaudio
from stable_audio_3 import StableAudioModel
def main():
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--wav", help="init audio (beatbox) WAV; omit for pure text-to-audio")
ap.add_argument("--prompt", default="brutal metal blast beat drums, aggressive double bass, punchy triggered kick and snare")
ap.add_argument("--out", default="out.wav")
ap.add_argument("--model", default="small-sfx")
ap.add_argument("--steps", type=int, default=8)
ap.add_argument("--cfg", type=float, default=1.0)
ap.add_argument("--noise", type=float, default=0.9)
ap.add_argument("--seed", type=int, default=-1)
ap.add_argument("--sampler", default="pingpong")
ap.add_argument("--duration", type=float, default=None)
args = ap.parse_args()
device = "cpu"
print(f"Loading model {args.model} on {device} ...")
model = StableAudioModel.from_pretrained(args.model, device=device, model_half=False)
init_audio = None
if args.wav:
waveform, sr = torchaudio.load(args.wav)
clip_len = waveform.shape[-1] / sr
init_audio = (sr, waveform)
duration = args.duration or clip_len
print(f" init audio {args.wav}: {clip_len:.1f}s, sr={sr}")
else:
duration = args.duration or 7.0
print(" pure text-to-audio (no init audio)")
print(f" steps={args.steps}, cfg={args.cfg}, sampler={args.sampler}, "
f"noise={args.noise}, seed={args.seed}, duration={duration}s")
audio = model.generate(
prompt=args.prompt,
duration=duration,
steps=args.steps,
cfg_scale=args.cfg,
seed=args.seed,
sampler_type=args.sampler,
apg_scale=1.0,
init_audio=init_audio,
init_noise_level=args.noise,
)
out = audio.squeeze(0).cpu()
torchaudio.save(args.out, out, 44100)
print(f"Wrote {args.out}")
if __name__ == "__main__":
main()
|