import copy import json import os import os.path as osp import random from argparse import ArgumentParser from datetime import datetime import gradio as gr import moviepy.editor as mpy import numpy as np import torch from huggingface_hub import hf_hub_download from omegaconf import OmegaConf from PIL import Image, ImageDraw, ImageFont from animatediff.pipelines import I2VPipeline from animatediff.utils.util import RANGE_LIST, save_videos_grid sample_idx = 0 def convert_gif_to_mp4(gif_path, mp4_path): clip = mpy.VideoFileClip(gif_path) clip.write_videofile(mp4_path) def add_text(gif_file: str, mp4_file: str, text_index: int): image = Image.open(gif_file) frames = [] try: while True: frames.append(image.copy().convert('RGB')) image.seek(len(frames)) except EOFError: pass text = ['Earn More Money!', 'Happy New Year!', 'Bad Luck Go Away!', 'Happy New Year!'][text_index] size = [36, 36, 36, 36][text_index] for i, frame in enumerate(frames): font = ImageFont.truetype('zyhzx.ttf', size=size + i * 2) draw = ImageDraw.Draw(frame) text_width, text_height = draw.textsize(text, font=font) image_width, image_height = image.size x = (image_width - text_width) // 2 y = (image_height - text_height) - (image_height - text_height) // 8 draw.text((x, y), text, fill='red', font=font) frames[0].save(gif_file, save_all=True, append_images=frames[1:], loop=0) mp4_file = convert_gif_to_mp4(gif_file, mp4_file) return gif_file, mp4_file css = """ .toolbutton { margin-buttom: 0em 0em 0em 0em; max-width: 2.5em; min-width: 2.5em !important; height: 2.5em; } """ parser = ArgumentParser() parser.add_argument('--config', type=str, default='example/config/base.yaml') parser.add_argument('--server-name', type=str, default='0.0.0.0') parser.add_argument('--port', type=int, default=7860) parser.add_argument('--share', action='store_true') parser.add_argument('--local-debug', action='store_true') parser.add_argument('--save-path', default='samples') args = parser.parse_args() LOCAL_DEBUG = args.local_debug BASE_CONFIG = 'example/config/base.yaml' STYLE_CONFIG_LIST = { '3d_cartoon': './example/openxlab/3-3d.yaml', } # download models PIA_PATH = './models/PIA' VAE_PATH = './models/VAE' DreamBooth_LoRA_PATH = './models/DreamBooth_LoRA' def seed_everything(seed): import random import numpy as np torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) np.random.seed(seed % (2**32)) random.seed(seed) if not LOCAL_DEBUG: CACHE_PATH = './models' PIA_PATH = osp.join(CACHE_PATH, 'PIA') VAE_PATH = osp.join(CACHE_PATH, 'VAE') DreamBooth_LoRA_PATH = osp.join(CACHE_PATH, 'DreamBooth_LoRA') STABLE_DIFFUSION_PATH = osp.join(CACHE_PATH, 'StableDiffusion') os.makedirs(PIA_PATH, exist_ok=True) os.makedirs(VAE_PATH, exist_ok=True) os.makedirs(DreamBooth_LoRA_PATH, exist_ok=True) os.makedirs(STABLE_DIFFUSION_PATH, exist_ok=True) PIA_PATH = hf_hub_download(repo_id='Leoxing/PIA', filename='pia.ckpt', cache_dir=PIA_PATH) PIA_PATH = '/'.join(PIA_PATH.split('/')[:-1]) os.system('bash download_bashscripts/2-RcnzCartoon.sh') print(os.listdir(DreamBooth_LoRA_PATH)) # unet unet_full_path = hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='unet', filename='diffusion_pytorch_model.bin', cache_dir='models/StableDiffusion') STABLE_DIFFUSION_PATH = '/'.join(unet_full_path.split('/')[:-2]) hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='unet', filename='config.json', cache_dir='models/StableDiffusion') # vae hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='vae', filename='config.json', cache_dir='models/StableDiffusion') hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='vae', filename='diffusion_pytorch_model.bin', cache_dir='models/StableDiffusion') # text encoder hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='text_encoder', filename='config.json', cache_dir='models/StableDiffusion') hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='text_encoder', filename='pytorch_model.bin', cache_dir='models/StableDiffusion') # tokenizer hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='tokenizer', filename='merges.txt', cache_dir='models/StableDiffusion') hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='tokenizer', filename='special_tokens_map.json', cache_dir='models/StableDiffusion') hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='tokenizer', filename='tokenizer_config.json', cache_dir='models/StableDiffusion') hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='tokenizer', filename='vocab.json', cache_dir='models/StableDiffusion') # scheduler hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', subfolder='scheduler', filename='scheduler_config.json', cache_dir='models/StableDiffusion') # model index hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', filename='model_index.json', cache_dir='models/StableDiffusion') else: PIA_PATH = './models/PIA' VAE_PATH = './models/VAE' DreamBooth_LoRA_PATH = './models/DreamBooth_LoRA' STABLE_DIFFUSION_PATH = './models/StableDiffusion/sd15' def preprocess_img(img_np, max_size: int = 1024): ori_image = Image.fromarray(img_np).convert('RGB') width, height = ori_image.size short_edge = max(width, height) if short_edge > max_size: scale_factor = max_size / short_edge else: scale_factor = 1 width = int(width * scale_factor) height = int(height * scale_factor) ori_image = ori_image.resize((width, height)) if (width % 8 != 0) or (height % 8 != 0): in_width = (width // 8) * 8 in_height = (height // 8) * 8 else: in_width = width in_height = height in_image = ori_image in_image = ori_image.resize((in_width, in_height)) in_image_np = np.array(in_image) return in_image_np, in_height, in_width class AnimateController: def __init__(self): # config dirs self.basedir = os.getcwd() self.savedir = os.path.join( self.basedir, args.save_path, datetime.now().strftime("Gradio-%Y-%m-%dT%H-%M-%S")) self.savedir_sample = os.path.join(self.savedir, "sample") os.makedirs(self.savedir, exist_ok=True) self.inference_config = OmegaConf.load(args.config) self.style_configs = {k: OmegaConf.load( v) for k, v in STYLE_CONFIG_LIST.items()} self.pipeline_dict = self.load_model_list() def load_model_list(self): pipeline_dict = dict() for style, cfg in self.style_configs.items(): dreambooth_path = cfg.get('dreambooth', 'none') if dreambooth_path and dreambooth_path.upper() != 'NONE': dreambooth_path = osp.join( DreamBooth_LoRA_PATH, dreambooth_path) lora_path = cfg.get('lora', None) if lora_path is not None: lora_path = osp.join(DreamBooth_LoRA_PATH, lora_path) lora_alpha = cfg.get('lora_alpha', 0.0) vae_path = cfg.get('vae', None) if vae_path is not None: vae_path = osp.join(VAE_PATH, vae_path) pipeline_dict[style] = I2VPipeline.build_pipeline( self.inference_config, STABLE_DIFFUSION_PATH, unet_path=osp.join(PIA_PATH, 'pia.ckpt'), dreambooth_path=dreambooth_path, lora_path=lora_path, lora_alpha=lora_alpha, vae_path=vae_path, ip_adapter_path='h94/IP-Adapter', ip_adapter_scale=0.1) return pipeline_dict def fetch_default_n_prompt(self, style: str): cfg = self.style_configs[style] n_prompt = cfg.get('n_prompt', '') ip_adapter_scale = cfg.get('ip_adapter_scale', 0) gr.Info('Set default negative prompt and ip_adapter_scale.') print('Set default negative prompt and ip_adapter_scale.') return n_prompt, ip_adapter_scale def animate( self, init_img, motion_scale, prompt_textbox, negative_prompt_textbox, sample_step_slider, cfg_scale_slider, seed_textbox, ip_adapter_scale, style, progress=gr.Progress(), ): global sample_idx if init_img is None: gr.Warning('Please upload image or use example images.') if seed_textbox != -1 and seed_textbox != "": torch.manual_seed(int(seed_textbox)) seed = int(seed_textbox) else: seed = torch.initial_seed() generator = torch.Generator(device='cuda') generator.manual_seed(seed) seed_everything(seed) print(f'Seed: {seed}') pipeline = self.pipeline_dict[style] init_img, h, w = preprocess_img(init_img) print(f'img size: {h, w}') sample = pipeline( image=init_img, prompt=prompt_textbox, negative_prompt=negative_prompt_textbox, generator=generator, num_inference_steps=sample_step_slider, guidance_scale=cfg_scale_slider, width=w, height=h, video_length=16, mask_sim_template_idx=motion_scale - 1, ip_adapter_scale=ip_adapter_scale, progress_fn=progress, ).videos save_sample_path_mp4 = os.path.join( self.savedir_sample, f"{sample_idx}.mp4") save_sample_path_gif = os.path.join( self.savedir_sample, f"{sample_idx}.gif") save_videos_grid(sample, save_sample_path_mp4) save_videos_grid(sample, save_sample_path_gif) sample_config = { "prompt": prompt_textbox, "n_prompt": negative_prompt_textbox, "num_inference_steps": sample_step_slider, "guidance_scale": cfg_scale_slider, "width": w, "height": h, "seed": seed, "motion": motion_scale, } print(sample_config) json_str = json.dumps(sample_config, indent=4) with open(os.path.join(self.savedir, "logs.json"), "a") as f: f.write(json_str) f.write("\n\n") sample_idx += 1 return (save_sample_path_mp4, [save_sample_path_mp4, save_sample_path_gif]) def animate_example( self, init_img, motion_scale, prompt_textbox, negative_prompt_textbox, sample_step_slider, cfg_scale_slider, seed_textbox, ip_adapter_scale, style, with_text=False, text_idx=0, progress=gr.Progress(), ): print('init img', init_img) print('motion', motion_scale) print('prompt', prompt_textbox) print('sample step', sample_step_slider) print('ip-adapter', ip_adapter_scale) print('seed', seed_textbox) global sample_idx if init_img is None: print('Fetch example!!!!!!!!!!!') init_img = np.array(Image.open('__assets__/image_animation/zhening/zhening.jpeg')) gr.Info('Use example image for quick run.') if seed_textbox != -1 and seed_textbox != "": torch.manual_seed(int(seed_textbox)) seed = int(seed_textbox) else: seed = torch.initial_seed() generator = torch.Generator(device='cuda') generator.manual_seed(seed) seed_everything(seed) print(f'Seed: {seed}') pipeline = self.pipeline_dict[style] init_img, h, w = preprocess_img(init_img) print(f'img size: {h, w}') sample = pipeline( image=init_img, prompt=prompt_textbox, negative_prompt=negative_prompt_textbox, generator=generator, num_inference_steps=sample_step_slider, guidance_scale=cfg_scale_slider, width=w, height=h, video_length=16, mask_sim_template_idx=motion_scale - 1, ip_adapter_scale=ip_adapter_scale, progress_fn=progress, ).videos save_sample_path_mp4 = os.path.join( self.savedir_sample, f"{sample_idx}.mp4") save_sample_path_gif = os.path.join( self.savedir_sample, f"{sample_idx}.gif") save_videos_grid(sample, save_sample_path_mp4) save_videos_grid(sample, save_sample_path_gif) sample_config = { "prompt": prompt_textbox, "n_prompt": negative_prompt_textbox, "num_inference_steps": sample_step_slider, "guidance_scale": cfg_scale_slider, "width": w, "height": h, "seed": seed, "motion": motion_scale, } print(sample_config) json_str = json.dumps(sample_config, indent=4) with open(os.path.join(self.savedir, "logs.json"), "a") as f: f.write(json_str) f.write("\n\n") if with_text: add_text(save_sample_path_gif, save_sample_path_mp4, text_idx) sample_idx += 1 return (save_sample_path_mp4, [save_sample_path_mp4, save_sample_path_gif], seed, motion_scale, cfg_scale_slider, ) controller = AnimateController() def ui(): with gr.Blocks(css=css) as demo: # build state for default buttons default_motion = gr.State(value=1) default_prompt1 = gr.State( value='lift a red envelope, Chinese new year') default_prompt2 = gr.State( value='smiling, Chinese costume, Chinese new year') default_prompt3 = gr.State( value='angry, Chinese costume, Chinese new year') default_prompt4 = gr.State(value='sparklers, Chinese new year') default_n_prompt = gr.State(value='wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg') default_seed = gr.State(10201304011203481448) default_ip_adapter_scale = gr.State(0.2) default_style = gr.State('3d_cartoon') default_cfg = gr.State(7.5) default_1_idx = gr.State(0) default_2_idx = gr.State(1) default_3_idx = gr.State(2) default_4_idx = gr.State(3) gr.HTML( "
Your Personalized Image Animator
" "
via Plug-and-Play Modules in Text-to-Image Models
" ) with gr.Row(): gr.Markdown( "
Project Page  " # noqa "Paper  " "Code  " # noqa "Try More Style: Click here!
" # noqa ) with gr.Row(equal_height=False): with gr.Column(): with gr.Row(): init_img = gr.Image(label='Input Image') gr.Markdown('## Fast Try!') with gr.Row(): with gr.Column(scale=1, min_width=50): default_1 = gr.Button( '🧧', variant='primary', size='sm') with gr.Column(scale=1, min_width=50): default_2 = gr.Button( '🤗', variant='primary', size='sm') with gr.Column(scale=1, min_width=50): default_3 = gr.Button( '😡', variant='primary', size='sm') with gr.Column(scale=1, min_width=50): default_4 = gr.Button( '🧨', variant='primary', size='sm') with gr.Column(scale=1.5, min_width=150): with_wishes = gr.Checkbox(label='With Wishes✨') # style_dropdown = gr.Dropdown(label='Style', choices=list( # STYLE_CONFIG_LIST.keys()), value=list(STYLE_CONFIG_LIST.keys())[0]) style_dropdown = gr.State('3d_cartoon') with gr.Row(): prompt_textbox = gr.Textbox(label="Prompt", lines=1) gift_button = gr.Button( value='🎁', elem_classes='toolbutton' ) def append_gift(prompt): rand = random.randint(0, 2) if rand == 1: prompt = prompt + 'wearing santa hats' elif rand == 2: prompt = prompt + 'lift a Christmas gift' else: prompt = prompt + 'in Christmas suit, lift a Christmas gift' gr.Info('Merry Christmas! Add magic to your prompt!') return prompt gift_button.click( fn=append_gift, inputs=[prompt_textbox], outputs=[prompt_textbox], ) motion_scale_silder = gr.Slider( label='Motion Scale (Larger value means larger motion but less identity consistency)', value=1, step=1, minimum=1, maximum=len(RANGE_LIST)) with gr.Accordion('Advance Options', open=False): negative_prompt_textbox = gr.Textbox( value=controller.fetch_default_n_prompt( list(STYLE_CONFIG_LIST.keys())[0])[0], label="Negative prompt", lines=2) sample_step_slider = gr.Slider( label="Sampling steps", value=25, minimum=10, maximum=100, step=1) cfg_scale_slider = gr.Slider( label="CFG Scale", value=7.5, minimum=0, maximum=20) ip_adapter_scale = gr.Slider( label='IP-Apdater Scale', value=controller.fetch_default_n_prompt( list(STYLE_CONFIG_LIST.keys())[0])[1], minimum=0, maximum=1) with gr.Row(): seed_textbox = gr.Textbox(label="Seed", value=-1) seed_button = gr.Button( value="\U0001F3B2", elem_classes="toolbutton") seed_button.click( fn=lambda x: random.randint(1, 1e8), outputs=[seed_textbox], queue=False ) generate_button = gr.Button( value="Generate", variant='primary') with gr.Column(): result_video = gr.Video( label="Generated Animation", interactive=False) with gr.Row(): download = gr.Files( file_types=['gif', 'mp4'], label='Donwload Output') generate_button.click( fn=controller.animate, inputs=[ init_img, motion_scale_silder, prompt_textbox, negative_prompt_textbox, sample_step_slider, cfg_scale_slider, seed_textbox, ip_adapter_scale, style_dropdown, ], outputs=[result_video, download]) default_1.click( fn=controller.animate_example, inputs=[ init_img, default_motion, default_prompt1, default_n_prompt, sample_step_slider, default_cfg, default_seed, default_ip_adapter_scale, default_style, with_wishes, default_1_idx, ], outputs=[ result_video, download, default_seed, default_motion, default_cfg, ]) default_2.click( fn=controller.animate_example, inputs=[ init_img, default_motion, default_prompt2, default_n_prompt, sample_step_slider, default_cfg, default_seed, default_ip_adapter_scale, default_style, with_wishes, default_2_idx, ], outputs=[ result_video, download, default_seed, default_motion, default_cfg, ]) default_3.click( fn=controller.animate_example, inputs=[ init_img, default_motion, default_prompt3, default_n_prompt, sample_step_slider, default_cfg, default_seed, default_ip_adapter_scale, default_style, with_wishes, default_3_idx, ], outputs=[ result_video, download, default_seed, default_motion, default_cfg, ]) default_4.click( fn=controller.animate_example, inputs=[ init_img, default_motion, default_prompt4, default_n_prompt, sample_step_slider, default_cfg, default_seed, default_ip_adapter_scale, default_style, with_wishes, default_4_idx, ], outputs=[ result_video, download, default_seed, default_motion, default_cfg, ]) def create_example(input_list): return gr.Examples( examples=input_list, inputs=[ init_img, result_video, prompt_textbox, negative_prompt_textbox, style_dropdown, motion_scale_silder, ], ) gr.Markdown( '### Merry Christmas!' ) create_example( [ [ '__assets__/image_animation/yiming/yiming.jpeg', '__assets__/image_animation/yiming/yiming.mp4', '1boy in Christmas suit, lift a Christmas gift', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 2, ], [ '__assets__/image_animation/yanhong/yanhong.png', '__assets__/image_animation/yanhong/yanhong.mp4', '1girl lift a Christmas gift', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 2, ], ], ) with gr.Accordion('More Examples for Style Transfer', open=False): create_example([ [ '__assets__/image_animation/style_transfer/anya/anya.jpg', '__assets__/image_animation/style_transfer/anya/2.mp4', '1girl open mouth ', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 3, ], [ '__assets__/image_animation/magnitude/genshin/genshin.jpg', '__assets__/image_animation/magnitude/genshin/3.mp4', 'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 3, ], ]) with gr.Accordion('More Examples for Prompt Changing', open=False): create_example( [ [ '__assets__/image_animation/rcnz/harry.png', '__assets__/image_animation/rcnz/1.mp4', '1boy smiling', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 2 ], [ '__assets__/image_animation/rcnz/harry.png', '__assets__/image_animation/rcnz/2.mp4', '1boy playing magic fire', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 2 ], [ '__assets__/image_animation/rcnz/harry.png', '__assets__/image_animation/rcnz/3.mp4', '1boy is waving hands', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 2 ] ]) with gr.Accordion('Examples for Motion Magnitude', open=False): create_example( [ [ '__assets__/image_animation/magnitude/labrador.png', '__assets__/image_animation/magnitude/1.mp4', 'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 1, ], [ '__assets__/image_animation/magnitude/labrador.png', '__assets__/image_animation/magnitude/2.mp4', 'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 2, ], [ '__assets__/image_animation/magnitude/labrador.png', '__assets__/image_animation/magnitude/3.mp4', 'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k', 'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg', '3d_cartoon', 3, ] ]) return demo if __name__ == "__main__": demo = ui() demo.queue(max_size=10) demo.launch(server_name=args.server_name, server_port=args.port, share=args.share, max_threads=40, allowed_paths=['pia.png', 'samples'])