import os, argparse import sys import gradio as gr # from scripts.gradio.i2v_test_application import Image2Video sys.path.insert(1, os.path.join(sys.path[0], 'lvdm')) import spaces from lvdm.models.samplers.ddim import DDIMSampler import os import time from omegaconf import OmegaConf import torch from scripts.evaluation.funcs import load_model_checkpoint, save_videos, batch_ddim_sampling, get_latent_z from utils.utils import instantiate_from_config from huggingface_hub import hf_hub_download from einops import repeat import torchvision.transforms as transforms from pytorch_lightning import seed_everything from einops import rearrange from cldm.model import load_state_dict import cv2 import torch print("cuda available:", torch.cuda.is_available()) from huggingface_hub import snapshot_download import os def download_model(): REPO_ID = 'fbnnb/TC_sketch' filename_list = ['tc_sketch.pt'] tar_dir = './checkpoints/tooncrafter_1024_interp_sketch/' if not os.path.exists(tar_dir): os.makedirs(tar_dir) for filename in filename_list: local_file = os.path.join(tar_dir, filename) if not os.path.exists(local_file): hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir=tar_dir, local_dir_use_symlinks=False) print("downloaded") def get_latent_z_with_hidden_states(model, videos): b, c, t, h, w = videos.shape x = rearrange(videos, 'b c t h w -> (b t) c h w') encoder_posterior, hidden_states = model.first_stage_model.encode(x, return_hidden_states=True) hidden_states_first_last = [] ### use only the first and last hidden states for hid in hidden_states: hid = rearrange(hid, '(b t) c h w -> b c t h w', t=t) hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2) hidden_states_first_last.append(hid_new) z = model.get_first_stage_encoding(encoder_posterior).detach() z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t) return z, hidden_states_first_last def extract_frames(video_path): # 動画ファイルを読み込む cap = cv2.VideoCapture(video_path) frame_list = [] frame_num = 0 while True: # フレームを読み込む ret, frame = cap.read() if not ret: break # フレームをリストに追加 frame_list.append(frame) frame_num += 1 print("load video length:", len(frame_list)) # 動画ファイルを閉じる cap.release() return frame_list resolution = '576_1024' resolution = (576, 1024) download_model() print("after download model") result_dir = "./results/" if not os.path.exists(result_dir): os.mkdir(result_dir) #ToonCrafterModel ckpt_path='checkpoints/tooncrafter_1024_interp_sketch/tc_sketch.pt' config_file='configs/inference_1024_v1.0.yaml' config = OmegaConf.load(config_file) model_config = config.pop("model", OmegaConf.create()) model_config['params']['unet_config']['params']['use_checkpoint']=False model = instantiate_from_config(model_config).cuda() assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!" model = load_model_checkpoint(model, ckpt_path) model.eval() # cn_model.load_state_dict(load_state_dict(cn_ckpt_path, location='cpu')) # cn_model.eval() # model.control_model = cn_model # model_list.append(model) save_fps = 8 print("resolution:", resolution) print("init done.") def transpose_if_needed(tensor): h = tensor.shape[-2] w = tensor.shape[-1] if h > w: tensor = tensor.permute(0, 2, 1) return tensor def untranspose(tensor): ndim = tensor.ndim return tensor.transpose(ndim-1, ndim-2) # [i2v_input_image, i2v_input_text, i2v_input_image, i2v_input_image2, i2v_steps, i2v_eta, i2v_motion, i2v_seed], @spaces.GPU(duration=200) def get_image(image1, prompt, image2, dim_steps=50, ddim_eta=1., fs=None, seed=123, \ unconditional_guidance_scale=1.0, cfg_img=None, text_input=False, multiple_cond_cfg=False, \ loop=False, interp=False, timestep_spacing='uniform', guidance_rescale=0.0, noise_shape=[72, 108], n_samples=1, **kwargs): with torch.no_grad(): seed_everything(seed) video_size = (576, 1024) transform = transforms.Compose([ transforms.Resize(min(video_size)), transforms.CenterCrop(video_size), # transforms.ToTensor(), # transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)) ]) image1 = torch.from_numpy(image1).permute(2, 0, 1).float().cuda() input_h, input_w = image1.shape[1:] image1 = (image1 / 255. - 0.5) * 2 image2 = torch.from_numpy(image2).permute(2, 0, 1).float().cuda() input_h, input_w = image2.shape[1:] image2 = (image2 / 255. - 0.5) * 2 # image1 = Image.open(file_list[2*idx]).convert('RGB') image_tensor1 = transform(image1).unsqueeze(1) # [c,1,h,w] # image2 = Image.open(file_list[2*idx+1]).convert('RGB') image_tensor2 = transform(image2).unsqueeze(1) # [c,1,h,w] frame_tensor1 = repeat(image_tensor1, 'c t h w -> c (repeat t) h w', repeat=8) frame_tensor2 = repeat(image_tensor2, 'c t h w -> c (repeat t) h w', repeat=8) videos = torch.cat([frame_tensor1, frame_tensor2], dim=1).unsqueeze(0) # frame_tensor = torch.cat([frame_tensor1, frame_tensor1], dim=1) # _, filename = os.path.split(file_list[idx*2]) global model model.cuda() ddim_sampler = DDIMSampler(model) if not multiple_cond_cfg else DDIMSampler_multicond(model) batch_size = 1 fs = torch.tensor([fs], dtype=torch.long, device=model.device) if not text_input: prompts = [""]*batch_size img = videos[:,:,0] #bchw img_emb = model.embedder(img) ## blc img_emb = model.image_proj_model(img_emb) cond_emb = model.get_learned_conditioning(prompts) cond = {"c_crossattn": [torch.cat([cond_emb,img_emb], dim=1)]} if model.model.conditioning_key == 'hybrid': z, hs = get_latent_z_with_hidden_states(model, videos) # b c t h w if loop or interp: img_cat_cond = torch.zeros_like(z) img_cat_cond[:,:,0,:,:] = z[:,:,0,:,:] img_cat_cond[:,:,-1,:,:] = z[:,:,-1,:,:] else: img_cat_cond = z[:,:,:1,:,:] img_cat_cond = repeat(img_cat_cond, 'b c t h w -> b c (repeat t) h w', repeat=z.shape[2]) cond["c_concat"] = [img_cat_cond] # b c 1 h w if unconditional_guidance_scale != 1.0: if model.uncond_type == "empty_seq": prompts = batch_size * [""] uc_emb = model.get_learned_conditioning(prompts) elif model.uncond_type == "zero_embed": uc_emb = torch.zeros_like(cond_emb) uc_img_emb = model.embedder(torch.zeros_like(img)) ## b l c uc_img_emb = model.image_proj_model(uc_img_emb) uc = {"c_crossattn": [torch.cat([uc_emb,uc_img_emb],dim=1)]} if model.model.conditioning_key == 'hybrid': uc["c_concat"] = [img_cat_cond] else: uc = None # # for i, h in enumerate(hs): # print("h:", h.shape) # hs[i] = hs[i][:,:,0,:,:].unsqueeze(2) additional_decode_kwargs = {'ref_context': hs} # additional_decode_kwargs = {'ref_context': None} ## we need one more unconditioning image=yes, text="" if multiple_cond_cfg and cfg_img != 1.0: uc_2 = {"c_crossattn": [torch.cat([uc_emb,img_emb],dim=1)]} if model.model.conditioning_key == 'hybrid': uc_2["c_concat"] = [img_cat_cond] kwargs.update({"unconditional_conditioning_img_nonetext": uc_2}) else: kwargs.update({"unconditional_conditioning_img_nonetext": None}) z0 = None cond_mask = None batch_variants = [] for _ in range(n_samples): if z0 is not None: cond_z0 = z0.clone() kwargs.update({"clean_cond": True}) else: cond_z0 = None if ddim_sampler is not None: samples, _ = ddim_sampler.sample(S=ddim_steps, conditioning=cond, batch_size=batch_size, shape=noise_shape, verbose=False, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=uc, eta=ddim_eta, cfg_img=cfg_img, mask=cond_mask, x0=cond_z0, fs=fs, timestep_spacing=timestep_spacing, guidance_rescale=guidance_rescale, **kwargs ) ## reconstruct from latent to pixel space batch_images = model.decode_first_stage(samples, **additional_decode_kwargs) index = list(range(samples.shape[2])) del index[1] del index[-2] samples = samples[:,:,index,:,:] ## reconstruct from latent to pixel space batch_images_middle = model.decode_first_stage(samples, **additional_decode_kwargs) batch_images[:,:,batch_images.shape[2]//2-1:batch_images.shape[2]//2+1] = batch_images_middle[:,:,batch_images.shape[2]//2-2:batch_images.shape[2]//2] batch_variants.append(batch_images) ## variants, batch, c, t, h, w batch_variants = torch.stack(batch_variants) # return batch_variants.permute(1, 0, 2, 3, 4, 5) prompt_str = prompt.replace("/", "_slash_") if "/" in prompt else prompt prompt_str = prompt_str.replace(" ", "_") if " " in prompt else prompt_str prompt_str=prompt_str[:40] if len(prompt_str) == 0: prompt_str = 'empty_prompt' result_dir = "./tmp/" save_videos(batch_image, result_dir, filenames=[prompt_str], fps=8) print(f"Saved in {prompt_str}. Time used: {(time.time() - start):.2f} seconds") model = model.cpu() saved_result_dir = os.path.join(result_dir, f"{prompt_str}.mp4") print("result saved to:", saved_result_dir) return saved_result_dir def dynamicrafter_demo(result_dir='./tmp/', res=1024): if res == 1024: resolution = '576_1024' css = """#input_img {max-width: 1024px !important} #output_vid {max-width: 1024px; max-height:576px}""" elif res == 512: resolution = '320_512' css = """#input_img {max-width: 512px !important} #output_vid {max-width: 512px; max-height: 320px} #input_img2 {max-width: 512px !important} #output_vid {max-width: 512px; max-height: 320px}""" elif res == 256: resolution = '256_256' css = """#input_img {max-width: 256px !important} #output_vid {max-width: 256px; max-height: 256px}""" else: raise NotImplementedError(f"Unsupported resolution: {res}") # image2video = Image2Video(result_dir, resolution=resolution) with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface: with gr.Tab(label='ToonCrafter_576x1024'): with gr.Column(): with gr.Row(): with gr.Column(): with gr.Row(): i2v_input_image = gr.Image(label="Input Image1",elem_id="input_img") # frame_guides = gr.Video(label="Input Guidance",elem_id="input_guidance", autoplay=True,show_share_button=True) with gr.Row(): i2v_input_text = gr.Text(label='Prompts') with gr.Row(): i2v_seed = gr.Slider(label='Random Seed', minimum=0, maximum=50000, step=1, value=123) i2v_eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='ETA', value=1.0, elem_id="i2v_eta") i2v_cfg_scale = gr.Slider(minimum=1.0, maximum=15.0, step=0.5, label='CFG Scale', value=7.5, elem_id="i2v_cfg_scale") with gr.Row(): i2v_steps = gr.Slider(minimum=1, maximum=60, step=1, elem_id="i2v_steps", label="Sampling steps", value=50) i2v_motion = gr.Slider(minimum=5, maximum=30, step=1, elem_id="i2v_motion", label="FPS", value=10) control_scale = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, elem_id="i2v_ctrl_scale", label="control_scale", value=0.6) i2v_end_btn = gr.Button("Generate") with gr.Column(): with gr.Row(): i2v_input_image2 = gr.Image(label="Input Image 2",elem_id="input_img2") with gr.Row(): i2v_output_video = gr.Video(label="Generated Video",elem_id="output_vid",autoplay=True,show_share_button=True) # s(model, prompts, image1, image2, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1., \ # unconditional_guidance_scale=1.0, cfg_img=None, fs=None, text_input=False, multiple_cond_cfg=False, \ # loop=False, interp=False, timestep_spacing='uniform', guidance_rescale=0.0, **kwargs): # gr.Examples(examples=i2v_examples_interp_1024, # inputs=[i2v_input_image, i2v_input_text, i2v_input_image, i2v_input_image2, [72, 108], 1, i2v_steps, i2v_eta, 1.0, None, i2v_motion, i2v_seed], # outputs=[i2v_output_video], # fn = get_image, # cache_examples=False, # ) img_size = [72, 108] i2v_end_btn.click(inputs=[i2v_input_image, i2v_input_text, i2v_input_image2, i2v_steps, i2v_eta, i2v_motion, i2v_seed], outputs=[i2v_output_video], fn = get_image ) return dynamicrafter_iface def get_parser(): parser = argparse.ArgumentParser() return parser if __name__ == "__main__": parser = get_parser() args = parser.parse_args() result_dir = os.path.join('./', 'results') dynamicrafter_iface = dynamicrafter_demo(result_dir) dynamicrafter_iface.queue(max_size=12) print("launching...") dynamicrafter_iface.launch(max_threads=1, share=True) # dynamicrafter_iface.launch(server_name='0.0.0.0', server_port=12345) # dynamicrafter_iface.launch() # print("launched...")