import gradio as gr import torch from diffusers import StableDiffusionPipeline, DDIMScheduler from utils import video_to_frames, add_dict_to_yaml_file, save_video, seed_everything # from diffusers.utils import export_to_video from tokenflow_pnp import TokenFlow from preprocess_utils import * from tokenflow_utils import * # load sd model #device = torch.device("cuda" if torch.cuda.is_available() else "cpu") if torch.cuda.is_available(): device = "cuda" elif torch.backends.mps.is_available(): device = "mps" else: device = "cpu" model_id = "stabilityai/stable-diffusion-2-1-base" to = torch.float16 if self.device == 'cuda' else torch.float32 # components for the Preprocessor scheduler = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler") vae = AutoencoderKL.from_pretrained(model_id, subfolder="vae", revision="fp16", torch_dtype=to).to(device) tokenizer = CLIPTokenizer.from_pretrained(model_id, subfolder="tokenizer") text_encoder = CLIPTextModel.from_pretrained(model_id, subfolder="text_encoder", revision="fp16", torch_dtype=to).to(device) unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet", revision="fp16", torch_dtype=to).to(device) # pipe for TokenFlow tokenflow_pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=to).to(device) if device == "cuda": tokenflow_pipe.enable_xformers_memory_efficient_attention() def randomize_seed_fn(): seed = random.randint(0, np.iinfo(np.int32).max) return seed def reset_do_inversion(): return True def get_example(): case = [ [ 'examples/wolf.mp4', ], [ 'examples/woman-running.mp4', ], [ 'examples/cutting_bread.mp4', ], [ 'examples/running_dog.mp4', ] ] return case def prep(config): # timesteps to save if config["sd_version"] == '2.1': model_key = "stabilityai/stable-diffusion-2-1-base" elif config["sd_version"] == '2.0': model_key = "stabilityai/stable-diffusion-2-base" elif config["sd_version"] == '1.5' or config["sd_version"] == 'ControlNet': model_key = "runwayml/stable-diffusion-v1-5" elif config["sd_version"] == 'depth': model_key = "stabilityai/stable-diffusion-2-depth" toy_scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler") toy_scheduler.set_timesteps(config["save_steps"]) print("config[save_steps]", config["save_steps"]) timesteps_to_save, num_inference_steps = get_timesteps(toy_scheduler, num_inference_steps=config["save_steps"], strength=1.0, device=device) print("YOOOO timesteps to save", timesteps_to_save) # seed_everything(config["seed"]) if not config["frames"]: # original non demo setting save_path = os.path.join(config["save_dir"], f'sd_{config["sd_version"]}', Path(config["data_path"]).stem, f'steps_{config["steps"]}', f'nframes_{config["n_frames"]}') os.makedirs(os.path.join(save_path, f'latents'), exist_ok=True) add_dict_to_yaml_file(os.path.join(config["save_dir"], 'inversion_prompts.yaml'), Path(config["data_path"]).stem, config["inversion_prompt"]) # save inversion prompt in a txt file with open(os.path.join(save_path, 'inversion_prompt.txt'), 'w') as f: f.write(config["inversion_prompt"]) else: save_path = None model = Preprocess(device, config, vae=vae, text_encoder=text_encoder, scheduler=scheduler, tokenizer=tokenizer, unet=unet) print(type(model.config["batch_size"])) frames, latents, total_inverted_latents, rgb_reconstruction = model.extract_latents( num_steps=model.config["steps"], save_path=save_path, batch_size=model.config["batch_size"], timesteps_to_save=timesteps_to_save, inversion_prompt=model.config["inversion_prompt"], ) return frames, latents, total_inverted_latents, rgb_reconstruction def preprocess_and_invert(input_video, frames, latents, inverted_latents, seed, randomize_seed, do_inversion, # save_dir: str = "latents", steps, n_timesteps = 50, batch_size: int = 8, n_frames: int = 40, inversion_prompt:str = '', ): sd_version = "2.1" height = 512 weidth: int = 512 print("n timesteps", n_timesteps) if do_inversion or randomize_seed: preprocess_config = {} preprocess_config['H'] = height preprocess_config['W'] = weidth preprocess_config['save_dir'] = 'latents' preprocess_config['sd_version'] = sd_version preprocess_config['steps'] = steps preprocess_config['batch_size'] = batch_size preprocess_config['save_steps'] = int(n_timesteps) preprocess_config['n_frames'] = n_frames preprocess_config['seed'] = seed preprocess_config['inversion_prompt'] = inversion_prompt preprocess_config['frames'] = video_to_frames(input_video) preprocess_config['data_path'] = input_video.split(".")[0] if randomize_seed: seed = randomize_seed_fn() seed_everything(seed) frames, latents, total_inverted_latents, rgb_reconstruction = prep(preprocess_config) print(total_inverted_latents.keys()) print(len(total_inverted_latents.keys())) frames = gr.State(value=frames) latents = gr.State(value=latents) inverted_latents = gr.State(value=total_inverted_latents) do_inversion = False return frames, latents, inverted_latents, do_inversion def edit_with_pnp(input_video, frames, latents, inverted_latents, seed, randomize_seed, do_inversion, steps, prompt: str = "a marble sculpture of a woman running, Venus de Milo", # negative_prompt: str = "ugly, blurry, low res, unrealistic, unaesthetic", pnp_attn_t: float = 0.5, pnp_f_t: float = 0.8, batch_size: int = 8, #needs to be the same as for preprocess n_frames: int = 40,#needs to be the same as for preprocess n_timesteps: int = 50, gudiance_scale: float = 7.5, inversion_prompt: str = "", #needs to be the same as for preprocess n_fps: int = 10, progress=gr.Progress(track_tqdm=True) ): config = {} config["sd_version"] = "2.1" config["device"] = device config["n_timesteps"] = int(n_timesteps) config["n_frames"] = n_frames config["batch_size"] = batch_size config["guidance_scale"] = gudiance_scale config["prompt"] = prompt config["negative_prompt"] = "ugly, blurry, low res, unrealistic, unaesthetic", config["pnp_attn_t"] = pnp_attn_t config["pnp_f_t"] = pnp_f_t config["pnp_inversion_prompt"] = inversion_prompt if do_inversion: frames, latents, inverted_latents, do_inversion = preprocess_and_invert( input_video, frames, latents, inverted_latents, seed, randomize_seed, do_inversion, steps, n_timesteps, batch_size, n_frames, inversion_prompt) do_inversion = False if randomize_seed: seed = randomize_seed_fn() seed_everything(seed) editor = TokenFlow(config=config,pipe=tokenflow_pipe, frames=frames.value, inverted_latents=inverted_latents.value) edited_frames = editor.edit_video() save_video(edited_frames, 'tokenflow_PnP_fps_30.mp4', fps=n_fps) # path = export_to_video(edited_frames) return 'tokenflow_PnP_fps_30.mp4', frames, latents, inverted_latents, do_inversion ######## # demo # ######## intro = """