app.py

import gradio as gr
from diffusers import AudioLDMControlNetPipeline, ControlNetModel
from pretty_midi import PrettyMIDI
import torch

if torch.cuda.is_available():
    device = "cuda"
    torch_dtype = torch.float16
else:
    device = "cpu"
    torch_dtype = torch.float32

controlnet = ControlNetModel.from_pretrained("lauraibnz/midi-audioldm", torch_dtype=torch_dtype)
pipe = AudioLDMControlNetPipeline.from_pretrained("cvssp/audioldm-m-full", controlnet=controlnet, torch_dtype=torch_dtype)
pipe = pipe.to(device)

def predict(prompt, midi_file="test.mid", audio_length_in_s=10, num_inference_steps=20, controlnet_conditioning_scale=1.0):
    midi = PrettyMIDI(midi_file)
    audio = pipe(prompt, midi=midi, audio_length_in_s=audio_length_in_s, num_inference_steps=num_inference_steps, controlnet_conditioning_scale=controlnet_conditioning_scale)
    return (16000, audio.audios.T)

demo = gr.Interface(fn=predict, inputs="text", outputs="audio")
demo.launch()