app.py

import gradio as gr
import os
import cv2
import numpy as np
from PIL import Image
from moviepy.editor import *

import torch
import random
import requests
from io import BytesIO
from diffusers import StableDiffusionPipeline
from diffusers import DDIMScheduler
from utils import *
from inversion_utils import *
from torch import autocast, inference_mode
import re

def get_frames(video_in):
    frames = []
    #resize the video
    clip = VideoFileClip(video_in)
    
    #check fps
    if clip.fps > 30:
        print("vide rate is over 30, resetting to 30")
        clip_resized = clip.resize(height=512)
        clip_resized.write_videofile("video_resized.mp4", fps=30)
    else:
        print("video rate is OK")
        clip_resized = clip.resize(height=512)
        clip_resized.write_videofile("video_resized.mp4", fps=clip.fps)
    
    print("video resized to 512 height")
    
    # Opens the Video file with CV2
    cap= cv2.VideoCapture("video_resized.mp4")
    
    fps = cap.get(cv2.CAP_PROP_FPS)
    print("video fps: " + str(fps))
    i=0
    while(cap.isOpened()):
        ret, frame = cap.read()
        if ret == False:
            break
        cv2.imwrite('kang'+str(i)+'.jpg',frame)
        frames.append('kang'+str(i)+'.jpg')
        i+=1
    
    cap.release()
    cv2.destroyAllWindows()
    print("broke the video into frames")
    
    return frames, fps

def create_video(frames, fps):
    print("building video result")
    clip = ImageSequenceClip(frames, fps=fps)
    clip.write_videofile("_result.mp4", fps=fps)
    
    return "_result.mp4"

def randomize_seed_fn(seed, randomize_seed):
    if randomize_seed:
        seed = random.randint(0, np.iinfo(np.int32).max)
    torch.manual_seed(seed)
    return seed

def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):

  #  inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf, 
  #  based on the code in https://github.com/inbarhub/DDPM_inversion
   
  #  returns wt, zs, wts:
  #  wt - inverted latent
  #  wts - intermediate inverted latents
  #  zs - noise maps

  sd_pipe.scheduler.set_timesteps(num_diffusion_steps)

  # vae encode image
  with autocast("cuda"), inference_mode():
      w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215).float()

  # find Zs and wts - forward process
  wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=False, num_inference_steps=num_diffusion_steps)
  return zs, wts



def sample(zs, wts, prompt_tar="", skip=36, cfg_scale_tar=15, eta = 1):

    # reverse process (via Zs and wT)
    w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=False, zs=zs[skip:])
    
    # vae decode image
    with autocast("cuda"), inference_mode():
        x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
    if x0_dec.dim()<4:
        x0_dec = x0_dec[None,:,:,:]
    img = image_grid(x0_dec)
    return img

# load pipelines
sd_model_id = "runwayml/stable-diffusion-v1-5"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
sd_pipe = StableDiffusionPipeline.from_pretrained(sd_model_id).to(device)
sd_pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler")

def infer(video_in, do_inversion, wts, zs,
                            src_prompt,
                            tar_prompt, 
                            steps,
                            cfg_scale_src,
                            cfg_scale_tar,
                            skip, seed, randomize_seed):
    
    
    # 1. break video into frames and get FPS
    break_vid = get_frames(video_in)
    frames_list= break_vid[0]
    fps = break_vid[1]
    #n_frame = int(trim_value*fps)
    n_frame = len(frames_list)
    
    if n_frame >= len(frames_list):
        print("video is shorter than the cut value")
        n_frame = len(frames_list)
    
    # 2. prepare frames result arrays
    result_frames = []
    print("set stop frames to: " + str(n_frame))
    
    for i, image in enumerate(frames_list[0:int(n_frame)]):
        #mmpose_frame = get_mmpose_filter(i)

        #seed operations
        seed = randomize_seed_fn(seed, randomize_seed)
        do_inversion = True

        image = Image.open(image).convert("RGB")
        image = np.array(image)
        output_frame = edit(image,
                            do_inversion, wts, zs,
                            src_prompt,
                            tar_prompt, 
                            steps,
                            cfg_scale_src,
                            cfg_scale_tar,
                            skip,
                            seed,
                            randomize_seed 
                           )
        print(output_frame)
        do_inversion = False
        image = output_frame[0]
        #image = Image.fromarray(output_frame[0])
        image.save("_frame_" + str(i) + ".jpeg")
        result_frames.append("_frame_" + str(i) + ".jpeg")
        print("frame " + str(i) + "/" + str(n_frame) + ": done;")

    
    final_vid = create_video(result_frames, fps)

    
    return final_vid

def get_example():
    case = [
        [
            'Examples/gnochi_mirror.jpeg', 
            'Watercolor painting of a cat sitting next to a mirror',
            'Examples/gnochi_mirror_watercolor_painting.png', 
            '',
            100,
            3.5,
            36,
            15,
 
             ],
        [
            'Examples/source_an_old_man.png', 
            'A bronze statue of an old man',
            'Examples/ddpm_a_bronze_statue_of_an_old_man.png', 
            '',
            100,
            3.5,
            36,
            15,
 
             ],
        [
            'Examples/source_a_ceramic_vase_with_yellow_flowers.jpeg', 
            'A pink ceramic vase with a wheat bouquet',
            'Examples/ddpm_a_pink_ceramic_vase_with_a_wheat_bouquet.png', 
            '',
            100,
            3.5,
            36,
            15,
 
             ],

        [
            'Examples/source_a_model_on_a_runway.jpeg', 
            'A zebra on the runway',
            'Examples/ddpm_a_zebra_on_the_run_way.png', 
            '',
            100,
            3.5,
            36,
            15,
 
             ]
    
    
    ]
    return case







########
# demo #
########
                        
intro = """
<h1 style="font-weight: 1400; text-align: center; margin-bottom: 7px;">
   Edit Friendly DDPM Inversion
</h1>
<p style="font-size: 0.9rem; text-align: center; margin: 0rem; line-height: 1.2em; margin-top:1em">
Based on the work introduced in:
<a href="https://arxiv.org/abs/2304.06140" style="text-decoration: underline;" target="_blank">An Edit Friendly DDPM Noise Space:
Inversion and Manipulations </a> 
<p/>
<p style="font-size: 0.9rem; margin: 0rem; line-height: 1.2em; margin-top:1em">
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
<a href="https://huggingface.co/spaces/LinoyTsaban/edit_friendly_ddpm_inversion?duplicate=true">
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
<p/>"""
with gr.Blocks(css='style.css') as demo:
    
    def reset_do_inversion():
        do_inversion = True
        return do_inversion


    def edit(input_image,
            do_inversion, 
             wts, zs,
            src_prompt ="", 
            tar_prompt="",
            steps=100,
            cfg_scale_src = 3.5,
            cfg_scale_tar = 15,
            skip=36,
            seed = 0,
            randomize_seed  = True):

        x0 = load_512(input_image, device=device)
    
        if do_inversion or randomize_seed:
            zs_tensor, wts_tensor = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=cfg_scale_src)
            wts = gr.State(value=wts_tensor)
            zs = gr.State(value=zs_tensor)
            do_inversion = False
        
        output = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=cfg_scale_tar)    
        return output, wts, zs, do_inversion
    
    gr.HTML(intro)
    wts = gr.State()
    zs = gr.State()
    do_inversion = gr.State(value=True)
    with gr.Row():
        video_in = gr.Video(source="upload", type="filepath")
        #input_image = gr.Image(label="Input Image", interactive=True)
        #input_image.style(height=365, width=365)
        #output_image = gr.Image(label=f"Edited Image", interactive=False)
        #output_image.style(height=365, width=365)
        final_vid = gr.Video()
    
    with gr.Row():
        tar_prompt = gr.Textbox(lines=1, label="Describe your desired edited output", interactive=True)

    with gr.Row():
        with gr.Column(scale=1, min_width=100):
            edit_button = gr.Button("Run")



    with gr.Accordion("Advanced Options", open=False):
        with gr.Row():
            with gr.Column():
                #inversion
                src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True, placeholder="describe the original image")
                steps = gr.Number(value=100, precision=0, label="Num Diffusion Steps", interactive=True)
                cfg_scale_src = gr.Slider(minimum=1, maximum=15, value=3.5, label=f"Source Guidance Scale", interactive=True)
            with gr.Column():
                # reconstruction
                skip = gr.Slider(minimum=0, maximum=60, value=36, step = 1, label="Skip Steps", interactive=True)
                cfg_scale_tar = gr.Slider(minimum=7, maximum=18,value=15, label=f"Target Guidance Scale", interactive=True)
                seed = gr.Number(value=0, precision=0, label="Seed", interactive=True)
                randomize_seed = gr.Checkbox(label='Randomize seed', value=False)
            

    edit_button.click(
        fn=infer,
        inputs=[video_in,
                do_inversion, wts, zs, 
            src_prompt, 
            tar_prompt,
            steps,
            cfg_scale_src,
            cfg_scale_tar,
            skip,
            seed,randomize_seed 
        ],
        outputs=[final_vid],
    )

    video_in.change(
        fn = reset_do_inversion,
        outputs = [do_inversion]
    )

    src_prompt.change(
        fn = reset_do_inversion,
        outputs = [do_inversion]
    )


 



demo.queue()
demo.launch(share=False)