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 = """
Based on the work introduced in: An Edit Friendly DDPM Noise Space: Inversion and Manipulations
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.