app.py

import gradio as gr
import spaces

import os
import sys
import time
import subprocess
import shutil

import random
from omegaconf import OmegaConf
from moviepy.editor import VideoFileClip
from PIL import Image
import torch
import numpy as np

from black_box_image_edit.instructpix2pix import InstructPix2Pix
from prepare_video import crop_and_resize_video
from edit_image import infer_video

sys.path.insert(0, "i2vgen-xl")
from utils import load_ddim_latents_at_t
from pipelines.pipeline_i2vgen_xl import I2VGenXLPipeline
from run_group_ddim_inversion import ddim_inversion
from run_group_pnp_edit import init_pnp
from diffusers import DDIMInverseScheduler, DDIMScheduler
from diffusers.utils import load_image
import imageio
from transformers import pipeline

# Initialize the translation pipeline
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")

DEBUG_MODE = False

demo_examples = [
    ["./demo/Man Walking.mp4", "./demo/Man Walking/edited_first_frame/turn the man into darth vader.png", "남자를 다스 베이더로 변신시키기", 0.1, 0.1, 1.0],
    ["./demo/A kitten turning its head on a wooden floor.mp4", "./demo/A kitten turning its head on a wooden floor/edited_first_frame/A dog turning its head on a wooden floor.png", "나무 바닥에서 고개를 돌리는 강아지", 0.2, 0.2, 0.5],
    ["./demo/An Old Man Doing Exercises For The Body And Mind.mp4", "./demo/An Old Man Doing Exercises For The Body And Mind/edited_first_frame/jack ma.png", "몸과 마음을 위한 운동을 하는 남자", 0.8, 0.8, 1.0],
    ["./demo/Ballet.mp4", "./demo/Ballet/edited_first_frame/van gogh style.png", "반 고흐 스타일의 발레를 추는 소녀", 1.0, 1.0, 1.0],
    ["./demo/A Couple In A Public Display Of Affection.mp4", "./demo/A Couple In A Public Display Of Affection/edited_first_frame/Snowing.png", "공공장소에서 애정을 과시하는 커플, 눈이 내림", 0.3, 0.3, 1.0]
]


TEMP_DIR = "_demo_temp"


image_edit_model = InstructPix2Pix()

@torch.no_grad()
@spaces.GPU(duration=30)
def perform_edit(video_path, prompt, force_512=False, seed=42, negative_prompt=""):
    edited_image_path = infer_video(image_edit_model, 
                video_path, 
                output_dir=TEMP_DIR, 
                prompt=prompt, 
                prompt_type="instruct", 
                force_512=force_512, 
                seed=seed, 
                negative_prompt=negative_prompt,
                overwrite=True)
    return edited_image_path


# Set up default inversion config file
config = {
    # DDIM inversion
    "inverse_config": {
        "image_size": [512, 512],
        "n_frames": 16,
        "cfg": 1.0,
        "target_fps": 8,
        "ddim_inv_prompt": "",
        "prompt": "",
        "negative_prompt": "",
    },
    "pnp_config": {
        "random_ratio": 0.0,
        "target_fps": 8,
    },
}
config = OmegaConf.create(config)

# Initialize the I2VGenXL pipeline
pipe = I2VGenXLPipeline.from_pretrained(
    "ali-vilab/i2vgen-xl",
    torch_dtype=torch.float16,
    variant="fp16",
).to("cuda:0")

# Initialize the DDIM inverse scheduler
inverse_scheduler = DDIMInverseScheduler.from_pretrained(
        "ali-vilab/i2vgen-xl",
        subfolder="scheduler",
)
# Initialize the DDIM scheduler
ddim_scheduler = DDIMScheduler.from_pretrained(
        "ali-vilab/i2vgen-xl",
        subfolder="scheduler",
)

@torch.no_grad()
@spaces.GPU(duration=150)
def perform_anyv2v(
                    video_path, 
                    video_prompt, 
                    video_negative_prompt,
                    edited_first_frame_path, 
                    conv_inj, 
                    spatial_inj, 
                    temp_inj, 
                    num_inference_steps,
                    guidance_scale,
                    ddim_init_latents_t_idx,
                    ddim_inversion_steps,
                    seed,
                    ):

    tmp_dir = os.path.join(TEMP_DIR, "AnyV2V")
    if os.path.exists(tmp_dir):
        shutil.rmtree(tmp_dir)
    os.makedirs(tmp_dir)

    ddim_latents_path = os.path.join(tmp_dir, "ddim_latents")

    def read_frames(video_path):
        frames = []
        with imageio.get_reader(video_path) as reader:
            for i, frame in enumerate(reader):
                pil_image = Image.fromarray(frame)
                frames.append(pil_image)
        return frames
    frame_list = read_frames(str(video_path))

    config.inverse_config.image_size = list(frame_list[0].size)
    config.inverse_config.n_steps = ddim_inversion_steps
    config.inverse_config.n_frames = len(frame_list)
    config.inverse_config.output_dir = ddim_latents_path
    ddim_init_latents_t_idx = min(ddim_init_latents_t_idx, num_inference_steps - 1)

    # Step 1. DDIM Inversion
    first_frame = frame_list[0]

    generator = torch.Generator(device="cuda:0")
    generator = generator.manual_seed(seed)
    _ddim_latents = ddim_inversion(
        config.inverse_config,
        first_frame,
        frame_list,
        pipe,
        inverse_scheduler,
        generator,
    )

    # Step 2. DDIM Sampling + PnP feature and attention injection
    # Load the edited first frame
    edited_1st_frame = load_image(edited_first_frame_path).resize(
        config.inverse_config.image_size, resample=Image.Resampling.LANCZOS
    )
    # Load the initial latents at t
    ddim_scheduler.set_timesteps(num_inference_steps)
    print(f"ddim_scheduler.timesteps: {ddim_scheduler.timesteps}")
    ddim_latents_at_t = load_ddim_latents_at_t(
        ddim_scheduler.timesteps[ddim_init_latents_t_idx],
        ddim_latents_path=ddim_latents_path,
    )
    print(
        f"ddim_scheduler.timesteps[t_idx]: {ddim_scheduler.timesteps[ddim_init_latents_t_idx]}"
    )
    print(f"ddim_latents_at_t.shape: {ddim_latents_at_t.shape}")

    # Blend the latents
    random_latents = torch.randn_like(ddim_latents_at_t)
    print(
        f"Blending random_ratio (1 means random latent): {config.pnp_config.random_ratio}"
    )
    mixed_latents = (
        random_latents * config.pnp_config.random_ratio
        + ddim_latents_at_t * (1 - config.pnp_config.random_ratio)
    )

    # Init Pnp
    config.pnp_config.n_steps = num_inference_steps
    config.pnp_config.pnp_f_t = conv_inj
    config.pnp_config.pnp_spatial_attn_t = spatial_inj
    config.pnp_config.pnp_temp_attn_t = temp_inj
    config.pnp_config.ddim_init_latents_t_idx = ddim_init_latents_t_idx
    init_pnp(pipe, ddim_scheduler, config.pnp_config)
    # Edit video
    pipe.register_modules(scheduler=ddim_scheduler)

    edited_video = pipe.sample_with_pnp(
        prompt=video_prompt,
        image=edited_1st_frame,
        height=config.inverse_config.image_size[1],
        width=config.inverse_config.image_size[0],
        num_frames=config.inverse_config.n_frames,
        num_inference_steps=config.pnp_config.n_steps,
        guidance_scale=guidance_scale,
        negative_prompt=video_negative_prompt,
        target_fps=config.pnp_config.target_fps,
        latents=mixed_latents,
        generator=generator,
        return_dict=True,
        ddim_init_latents_t_idx=ddim_init_latents_t_idx,
        ddim_inv_latents_path=ddim_latents_path,
        ddim_inv_prompt=config.inverse_config.ddim_inv_prompt,
        ddim_inv_1st_frame=first_frame,
    ).frames[0]

    edited_video = [
        frame.resize(config.inverse_config.image_size, resample=Image.LANCZOS)
        for frame in edited_video
    ]

    def images_to_video(images, output_path, fps=24):
        writer = imageio.get_writer(output_path, fps=fps)

        for img in images:
            img_np = np.array(img)
            writer.append_data(img_np)

        writer.close()
    output_path = os.path.join(tmp_dir, "edited_video.mp4")
    images_to_video(
        edited_video, output_path, fps=config.pnp_config.target_fps
    )
    return output_path



def get_first_frame_as_pil(video_path):
    with VideoFileClip(video_path) as clip:
        # Extract the first frame (at t=0) as an array
        first_frame_array = clip.get_frame(0)
        # Convert the numpy array to a PIL Image
        first_frame_image = Image.fromarray(first_frame_array)
        return first_frame_image
        
def btn_preprocess_video_fn(video_path, width, height, start_time, end_time, center_crop, x_offset, y_offset, longest_to_width):
    def check_video(video_path):
        with VideoFileClip(video_path) as clip:
            if clip.duration == 2 and clip.fps == 8:
                return True
            else:
                return False

    if check_video(video_path) == False:
        processed_video_path = crop_and_resize_video(input_video_path=video_path, 
                                                    output_folder=TEMP_DIR,
                                                    clip_duration=2,
                                                    width=width, 
                                                    height=height, 
                                                    start_time=start_time, 
                                                    end_time=end_time, 
                                                    center_crop=center_crop, 
                                                    x_offset=x_offset, 
                                                    y_offset=y_offset, 
                                                    longest_to_width=longest_to_width)

        return processed_video_path
    else:
        return video_path

def btn_image_edit_fn(video_path, instruct_prompt, ie_force_512, ie_seed, ie_neg_prompt):
    """
    Generate an image based on the video and text input.
    """
    # Translate the instruction prompt from Korean to English
    translated_prompt = translator(instruct_prompt, src_lang="ko", tgt_lang="en")[0]['translation_text']

    if ie_seed < 0:
        ie_seed = int.from_bytes(os.urandom(2), "big")
    print(f"Using seed: {ie_seed}")

    edited_image_path = perform_edit(video_path=video_path, 
                                             prompt=translated_prompt,
                                             force_512=ie_force_512,
                                             seed=ie_seed,
                                             negative_prompt=ie_neg_prompt)
    return edited_image_path


def btn_infer_fn(video_path, 
                video_prompt, 
                video_negative_prompt,
                edited_first_frame_path, 
                conv_inj, 
                spatial_inj, 
                temp_inj, 
                num_inference_steps,
                guidance_scale,
                ddim_init_latents_t_idx,
                ddim_inversion_steps,
                seed,
                ):
    # Translate the video prompt from Korean to English
    translated_video_prompt = translator(video_prompt, src_lang="ko", tgt_lang="en")[0]['translation_text']

    if seed < 0:
        seed = int.from_bytes(os.urandom(2), "big")
    print(f"Using seed: {seed}")

    result_video_path = perform_anyv2v(video_path=video_path,
                                                        video_prompt=translated_video_prompt,
                                                        video_negative_prompt=video_negative_prompt,
                                                        edited_first_frame_path=edited_first_frame_path,
                                                        conv_inj=conv_inj,
                                                        spatial_inj=spatial_inj,
                                                        temp_inj=temp_inj,
                                                        num_inference_steps=num_inference_steps,
                                                        guidance_scale=guidance_scale,
                                                        ddim_init_latents_t_idx=ddim_init_latents_t_idx,
                                                        ddim_inversion_steps=ddim_inversion_steps,
                                                        seed=seed)

    return result_video_path

# Create the UI
#=====================================
with gr.Blocks() as demo:
    gr.Markdown("# text + video + image")

    with gr.Tabs():
        with gr.TabItem('text + video + image'):
            with gr.Row():
                with gr.Column():
                    video_raw = gr.Video(label="Raw Video Input")
                    btn_pv = gr.Button("Preprocess Video")
                    
                with gr.Column():
                    video_input = gr.Video(label="Preprocessed Video Input", interactive=False)
                with gr.Column():
                    advanced_settings_pv = gr.Accordion("Advanced Settings for Video Preprocessing", open=False)
                    with advanced_settings_pv:
                        with gr.Column():
                            pv_width = gr.Number(label="Width", value=512, minimum=1, maximum=4096)
                            pv_height = gr.Number(label="Height", value=512, minimum=1, maximum=4096)
                            pv_start_time = gr.Number(label="Start Time (End time - Start time must be = 2)", value=0, minimum=0)
                            pv_end_time = gr.Number(label="End Time (End time - Start time must be = 2)", value=2, minimum=0)
                            pv_center_crop = gr.Checkbox(label="Center Crop", value=True)
                            pv_x_offset = gr.Number(label="Horizontal Offset (-1 to 1)", value=0, minimum=-1, maximum=1)
                            pv_y_offset = gr.Number(label="Vertical Offset (-1 to 1)", value=0, minimum=-1, maximum=1)
                            pv_longest_to_width = gr.Checkbox(label="Resize Longest Dimension to Width")
                    
            gr.Markdown("# Image Editing Stage")
            with gr.Row():
                with gr.Column():
                    src_first_frame = gr.Image(label="First Frame", type="filepath", interactive=False)
                    image_instruct_prompt = gr.Textbox(label="Editing instruction prompt")
                    btn_image_edit = gr.Button("Edit the first frame")
                with gr.Column():
                    image_input_output = gr.Image(label="Edited Frame", type="filepath")
                with gr.Column():
                    advanced_settings_image_edit = gr.Accordion("Advanced Settings for Image Editing", open=True)
                    with advanced_settings_image_edit:
                        with gr.Column():
                            ie_neg_prompt = gr.Textbox(label="Negative Prompt", value="low res, blurry, watermark, jpeg artifacts")
                            ie_seed = gr.Number(label="Seed (-1 means random)", value=-1, minimum=-1, maximum=sys.maxsize)
                            ie_force_512 = gr.Checkbox(label="Force resize to 512x512 before feeding into the image editing model")

            gr.Markdown("# Video Editing Stage")
            with gr.Row():
                with gr.Column():
                    video_prompt = gr.Textbox(label="Video description prompt")
                    settings_anyv2v = gr.Accordion("Settings for AnyV2V")
                    with settings_anyv2v:
                        with gr.Column():
                            av_pnp_f_t = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.2, label="Convolutional injection (pnp_f_t)")
                            av_pnp_spatial_attn_t = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.2, label="Spatial Attention injection (pnp_spatial_attn_t)")
                            av_pnp_temp_attn_t = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label="Temporal Attention injection (pnp_temp_attn_t)")
                    btn_infer = gr.Button("Run Video Editing")
                with gr.Column():
                    video_output = gr.Video(label="Video Output")
                with gr.Column():
                    advanced_settings_anyv2v = gr.Accordion("Advanced Settings for AnyV2V", open=False)
                    with advanced_settings_anyv2v:
                        with gr.Column():
                            av_ddim_init_latents_t_idx = gr.Number(label="DDIM Initial Latents t Index", value=0, minimum=0)
                            av_ddim_inversion_steps = gr.Number(label="DDIM Inversion Steps", value=100, minimum=1)
                            av_num_inference_steps = gr.Number(label="Number of Inference Steps", value=100, minimum=1)
                            av_guidance_scale = gr.Number(label="Guidance Scale", value=9, minimum=0)
                            av_seed = gr.Number(label="Seed (-1 means random)", value=42, minimum=-1, maximum=sys.maxsize)
                            av_neg_prompt = gr.Textbox(label="Negative Prompt", value="Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms")

    examples = gr.Examples(examples=demo_examples, 
                           label="Examples (Just click on Video Editing button after loading them into the UI)",
                            inputs=[video_input, image_input_output, video_prompt, av_pnp_f_t, av_pnp_spatial_attn_t, av_pnp_temp_attn_t])
    gr.Markdown('Reference: You can find good source videos from https://www.pexels.com/videos/')

    btn_pv.click(
        btn_preprocess_video_fn,
        inputs=[video_raw, pv_width, pv_height, pv_start_time, pv_end_time, pv_center_crop, pv_x_offset, pv_y_offset, pv_longest_to_width],
        outputs=video_input
    )

    btn_image_edit.click(
        btn_image_edit_fn,
        inputs=[video_input, image_instruct_prompt, ie_force_512, ie_seed, ie_neg_prompt],
        outputs=image_input_output
    )
    
    btn_infer.click(
        btn_infer_fn,
        inputs=[video_input, 
                video_prompt, 
                av_neg_prompt,
                image_input_output, 
                av_pnp_f_t, 
                av_pnp_spatial_attn_t, 
                av_pnp_temp_attn_t,
                av_num_inference_steps,
                av_guidance_scale,
                av_ddim_init_latents_t_idx,
                av_ddim_inversion_steps,
                av_seed],
        outputs=video_output
    )

    video_input.change(fn=get_first_frame_as_pil, inputs=video_input, outputs=src_first_frame)

#=====================================

# Minimizing usage of GPU Resources
torch.set_grad_enabled(False)


demo.launch()