# !/usr/bin/env python # -*- coding: UTF-8 -*- import gc import logging import numpy as np import torch import os from PIL import ImageFont from huggingface_hub import hf_hub_download from diffusers import (StableDiffusionXLPipeline, DiffusionPipeline,EulerDiscreteScheduler, UNet2DConditionModel,UniPCMultistepScheduler, AutoencoderKL,) from transformers import CLIPVisionModelWithProjection from transformers import CLIPImageProcessor import folder_paths from comfy.model_management import cleanup_models from comfy.clip_vision import load as clip_load from comfy.model_management import total_vram from .utils.utils import get_comic from .utils.gradio_utils import character_to_dict from .utils.load_models_utils import load_models, get_instance_path from .model_loader_utils import (story_maker_loader,kolor_loader,phi2narry, extract_content_from_brackets,narry_list,remove_punctuation_from_strings,phi_list,center_crop_s,center_crop, narry_list_pil,setup_seed,find_directories, apply_style,get_scheduler,set_attention_processor,load_character_files_on_running, save_results,nomarl_upscale,process_generation,SAMPLER_NAMES,SCHEDULER_NAMES,lora_lightning_list,pre_checkpoint,get_easy_function) global total_count, attn_count, cur_step, mask1024, mask4096, attn_procs, unet global sa32, sa64 global write global height_s, width_s photomaker_dir=os.path.join(folder_paths.models_dir, "photomaker") device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu" MAX_SEED = np.iinfo(np.int32).max dir_path = os.path.dirname(os.path.abspath(__file__)) fonts_path = os.path.join(dir_path, "fonts") fonts_lists = os.listdir(fonts_path) base_pt = os.path.join(photomaker_dir,"pt") if not os.path.exists(base_pt): os.makedirs(base_pt) class Storydiffusion_Model_Loader: def __init__(self): pass @classmethod def INPUT_TYPES(cls): return { "required": { "character_prompt": ("STRING", {"multiline": True, "default": "[Taylor] a woman img, wearing a white T-shirt, blue loose hair.\n" "[Lecun] a man img,wearing a suit,black hair."}), "repo_id": ("STRING", {"default": ""}), "ckpt_name": (["none"]+folder_paths.get_filename_list("checkpoints"),), "vae_id":(["none"]+folder_paths.get_filename_list("vae"),), "character_weights": (["none"]+find_directories(base_pt),), "lora": (["none"] + folder_paths.get_filename_list("loras"),), "lora_scale": ("FLOAT", {"default": 0.8, "min": 0.1, "max": 1.0, "step": 0.1}), "controlnet_model": (["none"] + folder_paths.get_filename_list("controlnet"),), "clip_vision": (["none"] + folder_paths.get_filename_list("clip_vision"),), "trigger_words": ("STRING", {"default": "best quality"}), "sampeler_name": (SAMPLER_NAMES,), "scheduler": (SCHEDULER_NAMES,), "sa32_degree": ( "FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.1, "round": 0.01}), "sa64_degree": ( "FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.1, "round": 0.01}), "width": ("INT", {"default": 768, "min": 256, "max": 2048, "step": 32, "display": "number"}), "height": ("INT", {"default": 768, "min": 256, "max": 2048, "step": 32, "display": "number"}), "photomake_mode": (["v1", "v2"],), "easy_function":("STRING", {"default": ""}), }, "optional": {"image": ("IMAGE",), "condition_image": ("IMAGE",), "model": ("MODEL",), "clip":("CLIP",), "vae":("VAE",), }, } RETURN_TYPES = ("STORY_DICT", ) RETURN_NAMES = ("model", ) FUNCTION = "story_model_loader" CATEGORY = "Storydiffusion" def story_model_loader(self,character_prompt, repo_id, ckpt_name, vae_id, character_weights, lora, lora_scale,controlnet_model, clip_vision,trigger_words,sampeler_name, scheduler, sa32_degree, sa64_degree, width, height, photomake_mode, easy_function,**kwargs): cf_model=kwargs.get("model") clip = kwargs.get("clip") front_vae=kwargs.get("vae") scheduler_choice = get_scheduler(sampeler_name,scheduler) scheduler={"name":sampeler_name,"scheduler":scheduler} id_number=len(character_prompt.splitlines()) if id_number > 2: id_number=2 image = kwargs.get("image") if isinstance(image,torch.Tensor): #print(image.shape) batch_num,_,_,_=image.size() model_type="img2img" if batch_num!=id_number: raise "role prompt numbers don't match input image numbers...example:2 roles need 2 input images," else: model_type = "txt2img" image=None logging.info(f"Process using {id_number} roles,mode is {model_type}....") if controlnet_model=="none": controlnet_path=None else: controlnet_path=folder_paths.get_full_path("controlnet", controlnet_model) condition_image=kwargs.get("condition_image") photomaker_path = os.path.join(photomaker_dir, f"photomaker-{photomake_mode}.bin") photomake_mode_=photomake_mode # load model (auraface, NF4, save_model, kolor_face,flux_pulid_name,pulid,quantized_mode,story_maker,make_dual_only, clip_vision_path,char_files,ckpt_path,lora,lora_path,use_kolor,photomake_mode,use_flux,onnx_provider,low_vram)=get_easy_function( easy_function,clip_vision,character_weights,ckpt_name,lora,repo_id,photomake_mode) photomaker_path,face_ckpt,photomake_mode,pulid_ckpt,face_adapter,kolor_ip_path=pre_checkpoint( photomaker_path, photomake_mode, kolor_face, pulid, story_maker, clip_vision_path,use_kolor,model_type) if total_vram > 45000.0: aggressive_offload = False offload = False elif 17000.0=2] #改成单句中双方括号方法，利于MS组句，[A]... [B]...[C] prompts_dual = [prompt for prompt in prompts_origin if len(extract_content_from_brackets(prompt))>=2] positions_nc = [index for index, prompt in enumerate(prompts_origin) if"[NC]" in prompt] # 找到NC的位置 prompts_no_dual = [prompt for prompt in prompts_origin if not len(extract_content_from_brackets(prompt)) >= 2] prompts_no_nc_dual = [prompt for prompt in prompts_no_dual if "[NC]" not in prompt] if len(char_origin) == 2: positions_char_1 = [index for index, prompt in enumerate(prompts_origin) if char_origin[0] in prompt][ 0] # 获取角色出现的索引列表，并获取首次出现的位置 positions_char_2 = [index for index, prompt in enumerate(prompts_origin) if char_origin[1] in prompt][ 0] # 获取角色出现的索引列表，并获取首次出现的位置 cn_dict = {} # {0:[img]} story_maker if isinstance(control_image, torch.Tensor) and story_maker and controlnet_path and not make_dual_only: f1, _, _, _ = control_image.size() if f1!=len(prompts_origin): raise "if using story-maker controlnet,The number of input control-images and scene prompts should be consistent!! " else: if f1 == 1: cn_image_load = [nomarl_upscale(control_image, width, height)] else: img_list = list(torch.chunk(control_image, chunks=f1)) cn_image_load = [nomarl_upscale(img, width, height) for img in img_list] for index, img in enumerate(cn_image_load): cn_dict[index] = [img] for i in positions_dual: # dual to {dual_num:[None]} cn_dict[i] = [None] if model_type=="img2img": d1, _, _, _ = image.size() if d1 == 1: image_load = [nomarl_upscale(image, width, height)] else: img_list = list(torch.chunk(image, chunks=d1)) image_load = [nomarl_upscale(img, width, height) for img in img_list] gen = process_generation(pipe, image_load, model_type, steps, img_style, denoise_or_ip_sacle, style_strength_ratio, cfg, seed, id_length, character_prompt, negative_prompt, scene_prompts, width, height, load_chars, lora, trigger_words,photomake_mode,use_kolor,use_flux,make_dual_only, kolor_face,pulid,story_maker,input_id_emb_s_dict, input_id_img_s_dict,input_id_emb_un_dict, input_id_cloth_dict,guidance,condition_image,empty_emb_zero,use_cf,cf_scheduler,controlnet_path,controlnet_scale,cn_dict) else: if story_maker: raise "story maker only suppport img2img now" upload_images = None gen = process_generation(pipe, upload_images, model_type, steps, img_style, denoise_or_ip_sacle, style_strength_ratio, cfg, seed, id_length, character_prompt, negative_prompt, scene_prompts, width, height, load_chars, lora, trigger_words,photomake_mode,use_kolor,use_flux,make_dual_only,kolor_face, pulid,story_maker,input_id_emb_s_dict, input_id_img_s_dict,input_id_emb_un_dict, input_id_cloth_dict,guidance,condition_image,empty_emb_zero,use_cf,cf_scheduler,controlnet_path,controlnet_scale,cn_dict) for value in gen: print(type(value)) image_pil_list = phi_list(value) image_pil_list_ms = image_pil_list.copy() if save_character: print("saving character...") save_results(pipe.unet) if prompts_dual: if not clip_vision_path: raise "need a clip_vison weight." if use_flux or use_kolor: raise "flux or kolor don't support MS diffsion." if model_type == "img2img": image_a = image_load[0] image_b = image_load[1] else: image_a = image_pil_list_ms[positions_char_1] image_b = image_pil_list_ms[positions_char_2] if story_maker: print("start sampler dual prompt using story maker") if make_dual_only: del pipe cleanup_models(keep_clone_weights_loaded=False) gc.collect() torch.cuda.empty_cache() controlnet_path=None pipe=story_maker_loader(clip_load,clip_vision_path,dir_path,ckpt_path,face_adapter,UniPCMultistepScheduler,controlnet_path,low_vram) mask_image_1=input_id_img_s_dict[role_name_list[0]][0] mask_image_2=input_id_img_s_dict[role_name_list[1]][0] if photomake_mode=="v2": face_info_1 = input_id_emb_un_dict[role_name_list[0]][0] face_info_2 = input_id_emb_un_dict[role_name_list[1]][0] else: face_info_1 = input_id_emb_s_dict[role_name_list[0]][0] face_info_2 = input_id_emb_s_dict[role_name_list[1]][0] cloth_info_1=None cloth_info_2=None if isinstance(condition_image,torch.Tensor): cloth_info_1 = input_id_cloth_dict[role_name_list[0]][0] cloth_info_2 = input_id_cloth_dict[role_name_list[1]][0] prompts_dual, negative_prompt = apply_style( img_style, prompts_dual, negative_prompt ) # 添加Lora trigger add_trigger_words = " " + trigger_words+ " style " if lora: prompts_dual = remove_punctuation_from_strings(prompts_dual) if lora not in lora_lightning_list: # 加速lora不需要trigger prompts_dual = [item + add_trigger_words for item in prompts_dual] prompts_dual = [item.replace(char_origin[0], char_describe[0]) for item in prompts_dual if char_origin[0] in item] prompts_dual = [item.replace(char_origin[1], char_describe[1]) for item in prompts_dual if char_origin[1] in item] prompts_dual = [item.replace("[", " ", ).replace("]", " ", ) for item in prompts_dual] image_dual = [] if model_type=="txt2img": setup_seed(seed) generator = torch.Generator(device=device).manual_seed(seed) for i,prompt in enumerate(prompts_dual): output = pipe( image=image_a, mask_image=mask_image_1, face_info=face_info_1, # first person image_2=image_b, mask_image_2=mask_image_2, face_info_2=face_info_2, # second person prompt=prompt, negative_prompt=negative_prompt, ip_adapter_scale=denoise_or_ip_sacle, lora_scale=lora_scale, num_inference_steps=steps, guidance_scale=cfg, height=height, width=width, controlnet_conditioning_scale=controlnet_scale, generator=generator, cloth=cloth_info_1, cloth_2=cloth_info_2 ).images[0] image_dual.append(output) else: #using ms diffusion print("start sampler dual prompt using ms-diffusion") if controlnet_path: if not isinstance(control_image,torch.Tensor): raise "using controlnet need controlnet image input" if width != height: square = max(height, width) new_height, new_width = square, square image_a = center_crop(image_a) image_b = center_crop(image_b) else: new_width = width new_height = height del pipe cleanup_models(keep_clone_weights_loaded=False) gc.collect() torch.cuda.empty_cache() from .model_loader_utils import msdiffusion_main image_dual = msdiffusion_main(image_a, image_b, prompts_dual, new_width, new_height, steps, seed, img_style, char_describe, char_origin, negative_prompt, clip_vision_path, model_type, lora, lora_path, lora_scale, trigger_words, ckpt_path, repo_id, guidance, mask_threshold, start_step, controlnet_path, control_image, controlnet_scale, cfg, guidance_list, scheduler_choice) j = 0 for i in positions_dual: # 重新将双人场景插入原序列 if width != height: img = center_crop_s(image_dual[j], width, height) else: img = image_dual[j] image_pil_list.insert(int(i), img) j += 1 image_list = narry_list(image_pil_list) torch.cuda.empty_cache() else: image_list = narry_list(image_pil_list) image = torch.from_numpy(np.fromiter(image_list, np.dtype((np.float32, (height, width, 3))))) gc.collect() torch.cuda.empty_cache() return (image, scene_prompts,) class Comic_Type: @classmethod def INPUT_TYPES(s): return {"required": {"image": ("IMAGE",), "scene_prompts": ("STRING", {"multiline": True, "forceInput": True, "default": ""}), "fonts_list": (fonts_lists,), "text_size": ("INT", {"default": 40, "min": 1, "max": 100}), "comic_type": (["Four_Pannel", "Classic_Comic_Style"],), "split_lines": ("STRING", {"default": "；"}), }} RETURN_TYPES = ("IMAGE",) ETURN_NAMES = ("image",) FUNCTION = "comic_gen" CATEGORY = "Storydiffusion" def comic_gen(self, image, scene_prompts, fonts_list, text_size, comic_type, split_lines): result = [item for index, item in enumerate(image)] total_results = narry_list_pil(result) font_choice = os.path.join(dir_path, "fonts", fonts_list) captions = scene_prompts.splitlines() if len(captions) > 1: captions = [caption.replace("(", "").replace(")", "") if "(" or ")" in caption else caption for caption in captions] # del () captions = [caption.replace("[NC]", "") for caption in captions] captions = [caption.replace("]", "").replace("[", "") for caption in captions] captions = [ caption.split("#")[-1] if "#" in caption else caption for caption in captions ] else: prompt_array = scene_prompts.replace(split_lines, "\n") captions = prompt_array.splitlines() font = ImageFont.truetype(font_choice, text_size) images = ( get_comic(total_results, comic_type, captions=captions, font=font) + total_results ) images = phi2narry(images[0]) return (images,) class Pre_Translate_prompt: @classmethod def INPUT_TYPES(s): return {"required": {"scene_prompts": ("STRING", {"forceInput": True, "default": ""}), "keep_character_name": ("BOOLEAN", {"default": False},) }} RETURN_TYPES = ("STRING",) ETURN_NAMES = ("prompt_array",) FUNCTION = "translate_prompt" CATEGORY = "Storydiffusion" def translate_prompt(self, scene_prompts, keep_character_name): captions = scene_prompts.splitlines() if not keep_character_name: captions = [caption.split(")", 1)[-1] if ")" in caption else caption for caption in captions] # del character else: captions = [caption.replace("(", "").replace(")", "") if "(" or ")" in caption else caption for caption in captions] # del () captions = [caption.replace("[NC]", "") for caption in captions] if not keep_character_name: captions = [caption.split("]", 1)[-1] for caption in captions] else: captions = [caption.replace("]", "").replace("[", "") for caption in captions] captions = [ caption.split("#")[-1] if "#" in caption else caption for caption in captions ] scene_prompts = ''.join(captions) return (scene_prompts,) class Story_Easy_Function: @classmethod def INPUT_TYPES(s): return {"required": {"function_repo": ("STRING", {"default": ""}), "function_ckpt": (["none"] + folder_paths.get_filename_list("checkpoints"),), "function_cli": ("STRING", {"default": ""}), }} RETURN_TYPES = ("CLIP",) ETURN_NAMES = ("clip",) FUNCTION = "funcion_main" CATEGORY = "Storydiffusion" def funcion_main(self, function_repo, function_ckpt,function_cli): if "glm" in function_cli: pass #print("test") return (function_repo,) NODE_CLASS_MAPPINGS = { "Storydiffusion_Model_Loader": Storydiffusion_Model_Loader, "Storydiffusion_Sampler": Storydiffusion_Sampler, "Pre_Translate_prompt": Pre_Translate_prompt, "Comic_Type": Comic_Type, "Story_Easy_Function":Story_Easy_Function } NODE_DISPLAY_NAME_MAPPINGS = { "Storydiffusion_Model_Loader": "Storydiffusion_Model_Loader", "Storydiffusion_Sampler": "Storydiffusion_Sampler", "Pre_Translate_prompt": "Pre_Translate_prompt", "Comic_Type": "Comic_Type", "Story_Easy_Function":"Story_Easy_Function" }