Spaces:
Runtime error
Runtime error
app.py
Browse files
app.py
CHANGED
@@ -1,59 +1,32 @@
|
|
|
|
1 |
import torch
|
2 |
import torchaudio
|
3 |
from einops import rearrange
|
4 |
-
import gradio as gr
|
5 |
-
import spaces
|
6 |
-
import os
|
7 |
-
import uuid
|
8 |
-
|
9 |
-
# Importing the model-related functions
|
10 |
from stable_audio_tools import get_pretrained_model
|
11 |
from stable_audio_tools.inference.generation import generate_diffusion_cond
|
12 |
|
13 |
-
|
14 |
-
def load_model():
|
15 |
-
print("Loading model...")
|
16 |
-
model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
|
17 |
-
print("Model loaded successfully.")
|
18 |
-
return model, model_config
|
19 |
-
|
20 |
-
# Function to set up, generate, and process the audio
|
21 |
-
@spaces.GPU(duration=120) # Allocate GPU only when this function is called
|
22 |
-
def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
23 |
-
print(f"Prompt received: {prompt}")
|
24 |
-
print(f"Settings: Duration={seconds_total}s, Steps={steps}, CFG Scale={cfg_scale}")
|
25 |
-
|
26 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
27 |
-
print(f"Using device: {device}")
|
28 |
-
|
29 |
-
# Fetch the Hugging Face token from the environment variable
|
30 |
-
hf_token = os.getenv('HF_TOKEN')
|
31 |
-
print(f"Hugging Face token: {hf_token}")
|
32 |
-
|
33 |
-
# Use pre-loaded model and configuration
|
34 |
-
model, model_config = load_model()
|
35 |
-
sample_rate = model_config["sample_rate"]
|
36 |
-
sample_size = model_config["sample_size"]
|
37 |
|
38 |
-
|
|
|
|
|
|
|
39 |
|
40 |
-
|
41 |
-
print("Model moved to device.")
|
42 |
|
|
|
43 |
# Set up text and timing conditioning
|
44 |
conditioning = [{
|
45 |
-
"prompt": prompt,
|
46 |
-
"seconds_start": 0,
|
47 |
-
"seconds_total":
|
48 |
}]
|
49 |
-
print(f"Conditioning: {conditioning}")
|
50 |
|
51 |
# Generate stereo audio
|
52 |
-
print("Generating audio...")
|
53 |
output = generate_diffusion_cond(
|
54 |
model,
|
55 |
-
steps=
|
56 |
-
cfg_scale=
|
57 |
conditioning=conditioning,
|
58 |
sample_size=sample_size,
|
59 |
sigma_min=0.3,
|
@@ -61,86 +34,27 @@ def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
|
61 |
sampler_type="dpmpp-3m-sde",
|
62 |
device=device
|
63 |
)
|
64 |
-
print("Audio generated.")
|
65 |
|
66 |
# Rearrange audio batch to a single sequence
|
67 |
output = rearrange(output, "b d n -> d (b n)")
|
68 |
-
print("Audio rearranged.")
|
69 |
|
70 |
-
# Peak normalize, clip, convert to int16
|
71 |
output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
gr.Slider(0, 47, value=30, label="Duration in Seconds"),
|
91 |
-
gr.Slider(10, 150, value=100, step=10, label="Number of Diffusion Steps"),
|
92 |
-
gr.Slider(1, 15, value=7, step=0.1, label="CFG Scale")
|
93 |
-
],
|
94 |
-
outputs=gr.Audio(type="filepath", label="Generated Audio"),
|
95 |
-
title="Stable Audio Generator",
|
96 |
-
description="Generate variable-length stereo audio at 44.1kHz from text prompts using Stable Audio Open 1.0.",
|
97 |
-
examples=[
|
98 |
-
[
|
99 |
-
"Create a serene soundscape of a quiet beach at sunset.", # Text prompt
|
100 |
-
|
101 |
-
45, # Duration in Seconds
|
102 |
-
100, # Number of Diffusion Steps
|
103 |
-
10, # CFG Scale
|
104 |
-
],
|
105 |
-
[
|
106 |
-
"Generate an energetic and bustling city street scene with distant traffic and close conversations.", # Text prompt
|
107 |
-
|
108 |
-
30, # Duration in Seconds
|
109 |
-
120, # Number of Diffusion Steps
|
110 |
-
5, # CFG Scale
|
111 |
-
],
|
112 |
-
[
|
113 |
-
"Simulate a forest ambiance with birds chirping and wind rustling through the leaves.", # Text prompt
|
114 |
-
60, # Duration in Seconds
|
115 |
-
140, # Number of Diffusion Steps
|
116 |
-
7.5, # CFG Scale
|
117 |
-
],
|
118 |
-
[
|
119 |
-
"Recreate a gentle rainfall with distant thunder.", # Text prompt
|
120 |
-
|
121 |
-
35, # Duration in Seconds
|
122 |
-
110, # Number of Diffusion Steps
|
123 |
-
8, # CFG Scale
|
124 |
-
|
125 |
-
],
|
126 |
-
[
|
127 |
-
"Imagine a jazz cafe environment with soft music and ambient chatter.", # Text prompt
|
128 |
-
25, # Duration in Seconds
|
129 |
-
90, # Number of Diffusion Steps
|
130 |
-
6, # CFG Scale
|
131 |
-
|
132 |
-
],
|
133 |
-
["Rock beat played in a treated studio, session drumming on an acoustic kit.",
|
134 |
-
30, # Duration in Seconds
|
135 |
-
100, # Number of Diffusion Steps
|
136 |
-
7, # CFG Scale
|
137 |
-
|
138 |
-
]
|
139 |
-
])
|
140 |
-
|
141 |
-
|
142 |
-
# Pre-load the model to avoid multiprocessing issues
|
143 |
-
model, model_config = load_model()
|
144 |
-
|
145 |
-
# Launch the Interface
|
146 |
-
interface.launch()
|
|
|
1 |
+
import gradio as gr
|
2 |
import torch
|
3 |
import torchaudio
|
4 |
from einops import rearrange
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
from stable_audio_tools import get_pretrained_model
|
6 |
from stable_audio_tools.inference.generation import generate_diffusion_cond
|
7 |
|
8 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
+
# Download model
|
11 |
+
model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
|
12 |
+
sample_rate = model_config["sample_rate"]
|
13 |
+
sample_size = model_config["sample_size"]
|
14 |
|
15 |
+
model = model.to(device)
|
|
|
16 |
|
17 |
+
def generate_audio(prompt, bpm, duration):
|
18 |
# Set up text and timing conditioning
|
19 |
conditioning = [{
|
20 |
+
"prompt": f"{bpm} BPM {prompt}",
|
21 |
+
"seconds_start": 0,
|
22 |
+
"seconds_total": duration
|
23 |
}]
|
|
|
24 |
|
25 |
# Generate stereo audio
|
|
|
26 |
output = generate_diffusion_cond(
|
27 |
model,
|
28 |
+
steps=100,
|
29 |
+
cfg_scale=7,
|
30 |
conditioning=conditioning,
|
31 |
sample_size=sample_size,
|
32 |
sigma_min=0.3,
|
|
|
34 |
sampler_type="dpmpp-3m-sde",
|
35 |
device=device
|
36 |
)
|
|
|
37 |
|
38 |
# Rearrange audio batch to a single sequence
|
39 |
output = rearrange(output, "b d n -> d (b n)")
|
|
|
40 |
|
41 |
+
# Peak normalize, clip, convert to int16, and save to file
|
42 |
output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
43 |
+
|
44 |
+
return sample_rate, output
|
45 |
+
|
46 |
+
inputs = [
|
47 |
+
gr.inputs.Textbox(label="Prompt"),
|
48 |
+
gr.inputs.Number(label="BPM", default=128),
|
49 |
+
gr.inputs.Number(label="Duration (seconds)", default=30)
|
50 |
+
]
|
51 |
+
|
52 |
+
output = gr.outputs.Audio(type="numpy", label="Generated Audio")
|
53 |
+
|
54 |
+
gr.Interface(
|
55 |
+
fn=generate_audio,
|
56 |
+
inputs=inputs,
|
57 |
+
outputs=output,
|
58 |
+
title="Stable Audio Generation",
|
59 |
+
description="Generate audio using Stable Audio Open 1.0"
|
60 |
+
).launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|