#!/usr/bin/env python
from __future__ import annotations
import os
import gradio as gr
from inference_followyourpose import merge_config_then_run
import sys
sys.path.append('FollowYourPose')
# result = subprocess.run(['bash', './data/download.sh'], stdout=subprocess.PIPE)
import subprocess
zip_file = './example_video.zip'
output_dir = './data'
subprocess.run(['unzip', zip_file, '-d', output_dir])
current_dir = os.getcwd()
print("path is :", current_dir)
print("current_dir is :", os.listdir(current_dir))
print("dir is :", os.listdir(os.path.join(current_dir,'data')))
print("data/example_video is :", os.listdir(os.path.join(current_dir,'data/example_video')))
HF_TOKEN = os.getenv('HF_TOKEN')
pipe = merge_config_then_run()
with gr.Blocks(css='style.css') as demo:
gr.HTML(
"""
πΊπΊπΊ Follow Your Pose πππ
Pose-Guided Text-to-Video Generation using Pose-Free Videos
TL;DR: We tune 2D stable-diffusion to generate the character videos from pose and text description.
""")
gr.HTML("""
In order to run the demo successfully, we recommend the length of video is about 3~5 seconds.
The temporal crop offset and sampling stride are used to adjust the starting point and interval of video samples.
Due to the GPU limit of this demo, it currently generates 8-frame videos. For generating longer videos (e.g. 32 frames) shown on our webpage, we recommend trying our GitHub code on your own GPU.
You may duplicate the space and upgrade to GPU in settings for better performance and faster inference without waiting in the queue.
""")
with gr.Row():
with gr.Column():
with gr.Accordion('Input Video', open=True):
# user_input_video = gr.File(label='Input Source Video')
user_input_video = gr.Video(label='Input Source Video', source='upload', type='numpy', format="mp4", visible=True).style(height="auto")
video_type = gr.Dropdown(
label='The type of input video',
choices=[
"Raw Video",
"Skeleton Video"
], value="Raw Video")
with gr.Accordion('Temporal Crop offset and Sampling Stride', open=False):
n_sample_frame = gr.Slider(label='Number of Frames',
minimum=0,
maximum=32,
step=1,
value=8)
stride = gr.Slider(label='Temporal stride',
minimum=0,
maximum=20,
step=1,
value=1)
with gr.Accordion('Spatial Crop offset', open=False):
left_crop = gr.Number(label='Left crop',
value=0,
precision=0)
right_crop = gr.Number(label='Right crop',
value=0,
precision=0)
top_crop = gr.Number(label='Top crop',
value=0,
precision=0)
bottom_crop = gr.Number(label='Bottom crop',
value=0,
precision=0)
offset_list = [
left_crop,
right_crop,
top_crop,
bottom_crop,
]
ImageSequenceDataset_list = [
n_sample_frame,
stride
] + offset_list
with gr.Accordion('Text Prompt', open=True):
target_prompt = gr.Textbox(label='Target Prompt',
info='The simple background may achieve better results(e.g., "beach", "moon" prompt is better than "street" and "market")',
max_lines=1,
placeholder='Example: "Iron man on the beach"',
value='Iron man on the beach')
run_button = gr.Button('Generate')
with gr.Column():
result = gr.Video(label='Result')
# result.style(height=512, width=512)
with gr.Accordion('DDIM Parameters', open=True):
num_steps = gr.Slider(label='Number of Steps',
info='larger value has better editing capacity, but takes more time and memory.',
minimum=0,
maximum=50,
step=1,
value=50)
guidance_scale = gr.Slider(label='CFG Scale',
minimum=0,
maximum=50,
step=0.1,
value=12.0)
with gr.Row():
from example import style_example
examples = style_example
gr.Examples(examples=examples,
inputs = [
user_input_video,
target_prompt,
num_steps,
guidance_scale,
video_type,
*ImageSequenceDataset_list
],
outputs=result,
fn=pipe.run,
cache_examples=True,
)
inputs = [
user_input_video,
target_prompt,
num_steps,
guidance_scale,
video_type,
*ImageSequenceDataset_list
]
target_prompt.submit(fn=pipe.run, inputs=inputs, outputs=result)
run_button.click(fn=pipe.run, inputs=inputs, outputs=result)
demo.queue().launch()
# demo.queue().launch(share=False, server_name='0.0.0.0', server_port=80)