import gradio as gr
import spaces
import json
import re
from gradio_client import Client
from moviepy.editor import VideoFileClip
from moviepy.audio.AudioClip import AudioClip
def extract_audio(video_in):
input_video = video_in
output_audio = 'audio.wav'
# Open the video file and extract the audio
video_clip = VideoFileClip(input_video)
audio_clip = video_clip.audio
# Save the audio as a .wav file
audio_clip.write_audiofile(output_audio, fps=44100) # Use 44100 Hz as the sample rate for .wav files
print("Audio extraction complete.")
return 'audio.wav'
def get_caption(image_in):
kosmos2_client = Client("https://ydshieh-kosmos-2.hf.space/")
kosmos2_result = kosmos2_client.predict(
image_in, # str (filepath or URL to image) in 'Test Image' Image component
"Detailed", # str in 'Description Type' Radio component
fn_index=4
)
print(f"KOSMOS2 RETURNS: {kosmos2_result}")
with open(kosmos2_result[1], 'r') as f:
data = json.load(f)
reconstructed_sentence = []
for sublist in data:
reconstructed_sentence.append(sublist[0])
full_sentence = ' '.join(reconstructed_sentence)
#print(full_sentence)
# Find the pattern matching the expected format ("Describe this image in detail:" followed by optional space and then the rest)...
pattern = r'^Describe this image in detail:\s*(.*)$'
# Apply the regex pattern to extract the description text.
match = re.search(pattern, full_sentence)
if match:
description = match.group(1)
print(description)
else:
print("Unable to locate valid description.")
# Find the last occurrence of "."
#last_period_index = full_sentence.rfind('.')
# Truncate the string up to the last period
#truncated_caption = full_sentence[:last_period_index + 1]
# print(truncated_caption)
#print(f"\n—\nIMAGE CAPTION: {truncated_caption}")
return description
def get_caption_from_MD(image_in):
client = Client("https://vikhyatk-moondream1.hf.space/")
result = client.predict(
image_in, # filepath in 'image' Image component
"Describe precisely the image.", # str in 'Question' Textbox component
api_name="/answer_question"
)
print(result)
return result
def get_magnet(prompt):
client = Client("https://fffiloni-magnet.hf.space/")
result = client.predict(
"facebook/magnet-small-10secs", # Literal['facebook/magnet-small-10secs', 'facebook/magnet-medium-10secs', 'facebook/magnet-small-30secs', 'facebook/magnet-medium-30secs', 'facebook/audio-magnet-small', 'facebook/audio-magnet-medium'] in 'Model' Radio component
"", # str in 'Model Path (custom models)' Textbox component
prompt, # str in 'Input Text' Textbox component
3, # float in 'Temperature' Number component
0.9, # float in 'Top-p' Number component
10, # float in 'Max CFG coefficient' Number component
1, # float in 'Min CFG coefficient' Number component
20, # float in 'Decoding Steps (stage 1)' Number component
10, # float in 'Decoding Steps (stage 2)' Number component
10, # float in 'Decoding Steps (stage 3)' Number component
10, # float in 'Decoding Steps (stage 4)' Number component
"prod-stride1 (new!)", # Literal['max-nonoverlap', 'prod-stride1 (new!)'] in 'Span Scoring' Radio component
api_name="/predict_full"
)
print(result)
return result[1]
def get_audioldm(prompt):
client = Client("https://haoheliu-audioldm2-text2audio-text2music.hf.space/")
result = client.predict(
prompt, # str in 'Input text' Textbox component
"Low quality.", # str in 'Negative prompt' Textbox component
10, # int | float (numeric value between 5 and 15) in 'Duration (seconds)' Slider component
3.5, # int | float (numeric value between 0 and 7) in 'Guidance scale' Slider component
45, # int | float in 'Seed' Number component
3, # int | float (numeric value between 1 and 5) in 'Number waveforms to generate' Slider component
fn_index=1
)
print(result)
audio_result = extract_audio(result)
return audio_result
def get_riffusion(prompt):
client = Client("https://fffiloni-spectrogram-to-music.hf.space/--replicas/1qwjx/")
result = client.predict(
prompt, # str in 'Musical prompt' Textbox component
"", # str in 'Negative prompt' Textbox component
"", # filepath in 'parameter_4' Audio component
10, # float (numeric value between 5 and 10) in 'Duration in seconds' Slider component
api_name="/predict"
)
print(result)
return result[1]
import re
import torch
from transformers import pipeline
zephyr_model = "HuggingFaceH4/zephyr-7b-beta"
mixtral_model = "mistralai/Mixtral-8x7B-Instruct-v0.1"
pipe = pipeline("text-generation", model=zephyr_model, torch_dtype=torch.bfloat16, device_map="auto")
agent_maker_sys = f"""
You are an AI whose job is to help users create their own music which its genre will reflect the character or scene from an image described by users.
In particular, you need to respond succintly with few musical words, in a friendly tone, write a musical prompt for a music generation model.
For example, if a user says, "a picture of a man in a black suit and tie riding a black dragon", provide immediately a musical prompt corresponding to the image description.
Immediately STOP after that. It should be EXACTLY in this format:
"A grand orchestral arrangement with thunderous percussion, epic brass fanfares, and soaring strings, creating a cinematic atmosphere fit for a heroic battle"
"""
instruction = f"""
<|system|>
{agent_maker_sys}
<|user|>
"""
@spaces.GPU(enable_queue=True)
def get_musical_prompt(user_prompt):
prompt = f"{instruction.strip()}\n{user_prompt}"
outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
pattern = r'\<\|system\|\>(.*?)\<\|assistant\|\>'
cleaned_text = re.sub(pattern, '', outputs[0]["generated_text"], flags=re.DOTALL)
print(f"SUGGESTED Musical prompt: {cleaned_text}")
return cleaned_text.lstrip("\n")
def infer(image_in, chosen_model):
gr.Info("Getting image caption with Kosmos2...")
user_prompt = get_caption(image_in)
gr.Info("Building a musical prompt according to the image caption ...")
musical_prompt = get_musical_prompt(user_prompt)
if chosen_model == "MAGNet" :
gr.Info("Now calling MAGNet for music...")
music_o = get_magnet(musical_prompt)
elif chosen_model == "AudioLDM-2" :
gr.Info("Now calling AudioLDM-2 for music...")
music_o = get_magnet(musical_prompt)
elif chosen_model == "Riffusion" :
gr.Info("Now calling Riffusion for music...")
music_o = get_riffusion(musical_prompt)
return musical_prompt, music_o
demo_title = "Image to Music V2"
description = "Get music from a picture"
css = """
#col-container{
margin: 0 auto;
max-width: 980px;
text-align: left;
}
"""
with gr.Blocks(css=css) as demo:
with gr.Column(elem_id="col-container"):
gr.HTML(f"""
{demo_title}
{description}
""")
with gr.Row():
with gr.Column():
image_in = gr.Image(
label = "Image reference",
type = "filepath",
elem_id = "image-in"
)
chosen_model = gr.Dropdown(
label = "Choose a model",
choices = [
"MAGNet",
"AudioLDM-2",
"Riffusion"
],
value = "MAGNet"
)
submit_btn = gr.Button("Make music from my pic !")
with gr.Column():
caption = gr.Textbox(
label = "Inspirational musical prompt",
max_lines = 3
)
result = gr.Audio(
label = "Music"
)
with gr.Column():
gr.Examples(
examples = [
["examples/monalisa.png", "MAGNet"],
["examples/santa.png", "MAGNet"],
["examples/ocean_poet.jpeg", "MAGNet"],
["examples/winter_hiking.png", "MAGNet"],
["examples/teatime.jpeg", "MAGNet"],
["examples/news_experts.jpeg", "MAGNet"],
["examples/chicken_adobo.jpeg", "MAGNet"]
],
fn = infer,
inputs = [image_in, chosen_model],
outputs = [caption, result],
cache_examples = False
)
submit_btn.click(
fn = infer,
inputs = [
image_in,
chosen_model
],
outputs =[
caption,
result
]
)
demo.queue(max_size=16).launch(show_api=False)