diff --git a/MS_Diffusion.py b/MS_Diffusion.py index 45df7b1..ba0a4bd 100644 --- a/MS_Diffusion.py +++ b/MS_Diffusion.py @@ -4,17 +4,16 @@ import numpy as np import torch import os -import random import sys import re import yaml from PIL import Image from huggingface_hub import hf_hub_download -from diffusers import (StableDiffusionXLPipeline, DDIMScheduler, ControlNetModel, +from diffusers import (StableDiffusionXLPipeline, DDIMScheduler, ControlNetModel, KDPM2AncestralDiscreteScheduler, LMSDiscreteScheduler, DPMSolverMultistepScheduler, DPMSolverSinglestepScheduler, EulerDiscreteScheduler, HeunDiscreteScheduler, - KDPM2DiscreteScheduler, + KDPM2DiscreteScheduler, EulerAncestralDiscreteScheduler, UniPCMultistepScheduler, AutoencoderKL, StableDiffusionXLControlNetPipeline, DDPMScheduler, LCMScheduler) from transformers import CLIPVisionModelWithProjection @@ -25,22 +24,16 @@ from .msdiffusion.utils import get_phrase_idx, get_eot_idx from diffusers.schedulers.scheduling_ddim import DDIMScheduler import folder_paths from comfy.utils import common_upscale - - -device = ( - "cuda" - if torch.cuda.is_available() - else "mps" if torch.backends.mps.is_available() else "cpu" -) +from omegaconf import OmegaConf MAX_SEED = np.iinfo(np.int32).max dir_path = os.path.dirname(os.path.abspath(__file__)) path_dir = os.path.dirname(dir_path) file_path = os.path.dirname(path_dir) - -controlnet_list=["controlnet-openpose-sdxl-1.0","controlnet-zoe-depth-sdxl-1.0","controlnet-scribble-sdxl-1.0","controlnet-tile-sdxl-1.0","controlnet-depth-sdxl-1.0","controlnet-canny-sdxl-1.0","MistoLine","sdxl-controlnet-seg"] - +controlnet_list = ["controlnet-openpose-sdxl-1.0", "controlnet-zoe-depth-sdxl-1.0", "controlnet-scribble-sdxl-1.0", + "controlnet-tile-sdxl-1.0", "controlnet-depth-sdxl-1.0", "controlnet-canny-sdxl-1.0", "MistoLine", + "sdxl-controlnet-seg"] control_paths = [] for search_path in folder_paths.get_folder_paths("diffusers"): @@ -48,24 +41,13 @@ for search_path in folder_paths.get_folder_paths("diffusers"): for root, subdir, files in os.walk(search_path, followlinks=True): if "config.json" in files: control_paths.append(os.path.relpath(root, start=search_path)) - control_paths = [z for z in control_paths if z.split("\\")[-1] in controlnet_list or z in controlnet_list] + control_paths = [z for z in control_paths if + z.split("\\")[-1] in controlnet_list or z in controlnet_list] if control_paths: control_paths = ["none"] + [x for x in control_paths if x] else: - control_paths = ["none",] - - -clip_paths = [] -for search_path in folder_paths.get_folder_paths("clip_vision"): - if os.path.exists(search_path): - for root, subdir, files in os.walk(search_path, followlinks=True): - if "config.json" in files: - clip_paths.append(os.path.relpath(root, start=search_path)) -if clip_paths: - clip_paths = ["none"] + [x for x in clip_paths if x] -else: - clip_paths = ["none",] + control_paths = ["none", ] def get_instance_path(path): @@ -74,22 +56,26 @@ def get_instance_path(path): instance_path = instance_path.replace('\\', "/") return instance_path -loras_path = get_instance_path(os.path.join(dir_path,"config","lora.yaml")) + +loras_path = get_instance_path(os.path.join(dir_path, "config", "lora.yaml")) + + def get_lora_dict(): # 打开并读取YAML文件 with open(loras_path, 'r', encoding="UTF-8") as stream: try: # 解析YAML文件内容 data = yaml.safe_load(stream) - + # 此时 'data' 是一个Python字典，里面包含了YAML文件的所有数据 # print(data) return data - + except yaml.YAMLError as exc: # 如果在解析过程中发生了错误，打印异常信息 print(exc) + datas = get_lora_dict() lora_lightning_list = datas["lightning_xl_lora"] @@ -104,6 +90,7 @@ def phi2narry(img): img = torch.from_numpy(np.array(img).astype(np.float32) / 255.0).unsqueeze(0) return img + def nomarl_upscale(img_tensor, width, height): samples = img_tensor.movedim(-1, 1) img = common_upscale(samples, width, height, "nearest-exact", "center") @@ -116,23 +103,28 @@ def phi2tensor(img): image = torch.from_numpy(image)[None,] return image + def add_pil(list, list_add, num): new_list = list[:num] + list_add + list[num:] return new_list + def tensor_to_image(tensor): # tensor = tensor.cpu() image_np = tensor.squeeze().mul(255).clamp(0, 255).byte().numpy() image = Image.fromarray(image_np, mode='RGB') return image + # get fonts list def has_parentheses(s): return bool(re.search(r'\(.*?\)', s)) + def contains_brackets(s): return '[' in s or ']' in s + def narry_list(list_in): for i in range(len(list_in)): value = list_in[i] @@ -140,12 +132,14 @@ def narry_list(list_in): list_in[i] = modified_value return list_in + def phi_list(list_in): for i in range(len(list_in)): value = list_in[i] list_in[i] = value return list_in + def narry_list_pil(list_in): for i in range(len(list_in)): value = list_in[i] @@ -153,29 +147,24 @@ def narry_list_pil(list_in): list_in[i] = modified_value return list_in -def get_local_path(file_path, model_path,file): + +def get_local_path(file_path, model_path, file): path = os.path.join(file_path, "models", file, model_path) model_path = os.path.normpath(path) if sys.platform.startswith('win32'): model_path = model_path.replace('\\', "/") return model_path + def extract_content_between_brackets(text): # 正则表达式匹配多对括号内的内容 return re.findall(r'\((.*?)\)', text) + def extract_content_from_brackets(text): # 正则表达式匹配多对方括号内的内容 return re.findall(r'\[(.*?)\]', text) -def setup_seed(seed): - torch.manual_seed(seed) - if device == "cuda": - torch.cuda.manual_seed_all(seed) - np.random.seed(seed) - random.seed(seed) - torch.backends.cudnn.deterministic = True - def get_image_path_list(folder_name): image_basename_list = os.listdir(folder_name) @@ -184,6 +173,7 @@ def get_image_path_list(folder_name): ) return image_path_list + def get_scheduler(name): scheduler = False if name == "Euler": @@ -226,6 +216,7 @@ def get_scheduler(name): scheduler = UniPCMultistepScheduler() return scheduler + def nomarl_upscale_topil(img_tensor, width, height): samples = img_tensor.movedim(-1, 1) img = common_upscale(samples, width, height, "nearest-exact", "center") @@ -233,53 +224,62 @@ def nomarl_upscale_topil(img_tensor, width, height): img_pil = tensor_to_image(samples) return img_pil -def instance_path(path, repo,file): + +def instance_path(path, repo, file): if repo == "": if path == "none": repo = "none" else: - model_path = get_local_path(file_path, path,file) + model_path = get_local_path(file_path, path, file) repo = get_instance_path(model_path) return repo + class MSdiffusion_Model_Loader: def __init__(self): self.counters = {} - + @classmethod def INPUT_TYPES(cls): return { "required": { - "ckpt_name": (["none"]+folder_paths.get_filename_list("checkpoints"),), + "ckpt_name": (["none"] + folder_paths.get_filename_list("checkpoints"),), "repo_id": ("STRING", {"default": ""}), - "clip_vision_local":(clip_paths,), "clip_vision_repo": ("STRING", {"default": ""}), - "vae_id": (["none"]+folder_paths.get_filename_list("vae"),), + "vae_id": (["none"] + folder_paths.get_filename_list("vae"),), "controlnet_diff": (control_paths,), - "controlnet_repo":("STRING", {"default": ""}), + "controlnet_repo": ("STRING", {"default": ""}), "lora": (["none"] + folder_paths.get_filename_list("loras"),), "lora_scale": ("FLOAT", {"default": 0.8, "min": 0.1, "max": 1.0, "step": 0.1}), "trigger_words": ("STRING", {"default": "best quality"}), - "scheduler": (scheduler_list,),} + "scheduler": (scheduler_list,), }, + "optional": {"clip_vision": ("CLIP_VISION",), + } } - - - RETURN_TYPES = ("MODEL","MODEL","MODEL","MODEL", "STRING",) - RETURN_NAMES = ("pipe","ms_model","image_encoder", "image_processor","info",) + + RETURN_TYPES = ("MSDIF_DICT",) + RETURN_NAMES = ("model",) FUNCTION = "ms_model_loader" CATEGORY = "MSdiffusion" - - def ms_model_loader(self, ckpt_name,repo_id,clip_vision_local, clip_vision_repo, vae_id,controlnet_diff,controlnet_repo,lora, lora_scale, trigger_words,scheduler, - ): + + def ms_model_loader(self, ckpt_name, repo_id, clip_vision_repo, vae_id, controlnet_diff, + controlnet_repo, lora, lora_scale, trigger_words, scheduler, **kwargs): + if clip_vision_repo: + device = ("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu") + clip_vision_path = get_instance_path(clip_vision_repo) + image_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_vision_path).to(device, + dtype=torch.float16) + use_repo = True + else: + image_encoder = kwargs.get("clip_vision") + device = image_encoder.load_device + use_repo = False scheduler_choice = get_scheduler(scheduler) - controlnet_model=instance_path(controlnet_diff, controlnet_repo,"diffusers") - clip_vision=instance_path(clip_vision_local, clip_vision_repo,"clip_vision") - ckpt_path=folder_paths.get_full_path("checkpoints", ckpt_name) - - if clip_vision=="none": - raise "need clip_vision" - + controlnet_model = instance_path(controlnet_diff, controlnet_repo, "diffusers") + + ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name) + if lora != "none": lora_path = folder_paths.get_full_path("loras", lora) lora_path = get_instance_path(lora_path) @@ -289,18 +289,20 @@ class MSdiffusion_Model_Loader: lora = lora.split("/")[-1] if "\\" in lora: lora = lora.split("\\")[-1] - + # load SDXL pipeline if not repo_id: original_config_file = os.path.join(dir_path, "config", "sd_xl_base.yaml") if controlnet_model != "none": # 启用control controlnet = ControlNetModel.from_pretrained( - controlnet_model,use_safetensors=True, + controlnet_model, use_safetensors=True, torch_dtype=torch.float16 ) try: - pipe = StableDiffusionXLControlNetPipeline.from_single_file(ckpt_path,controlnet=controlnet,original_config=original_config_file,torch_dtype=torch.float16, ) + pipe = StableDiffusionXLControlNetPipeline.from_single_file(ckpt_path, controlnet=controlnet, + original_config=original_config_file, + torch_dtype=torch.float16, ) except: try: pipe = StableDiffusionXLControlNetPipeline.from_single_file(ckpt_path, controlnet=controlnet, @@ -318,7 +320,7 @@ class MSdiffusion_Model_Loader: ckpt_path, original_config_file=original_config_file, torch_dtype=torch.float16) except: raise "model error" - + else: if controlnet_model != "none": # 启用control try: @@ -343,9 +345,9 @@ class MSdiffusion_Model_Loader: pipe = StableDiffusionXLPipeline.from_pretrained( repo_id, torch_dtype=torch.float16, add_watermarker=False, ) - - if vae_id!="none": - vae_id=folder_paths.get_full_path("vae", vae_id) + + if vae_id != "none": + vae_id = folder_paths.get_full_path("vae", vae_id) pipe.vae = AutoencoderKL.from_single_file(vae_id, torch_dtype=torch.float16).to(device) if lora != "none": if lora in lora_lightning_list: @@ -353,15 +355,15 @@ class MSdiffusion_Model_Loader: pipe.fuse_lora() else: pipe.load_lora_weights(lora_path, adapter_name=trigger_words) - pipe.fuse_lora(adapter_names=[trigger_words,], lora_scale=lora_scale) - + pipe.fuse_lora(adapter_names=[trigger_words, ], lora_scale=lora_scale) + pipe.scheduler = scheduler_choice.from_config(pipe.scheduler.config) pipe.enable_xformers_memory_efficient_attention() pipe.enable_freeu(s1=0.6, s2=0.4, b1=1.1, b2=1.2) pipe.enable_vae_slicing() - if device != "mps": - pipe.enable_model_cpu_offload() - ms_dir = os.path.join(file_path, "models","photomaker") + # if device != "mps": + # pipe.enable_model_cpu_offload() + ms_dir = os.path.join(file_path, "models", "photomaker") photomaker_local_path = os.path.join(ms_dir, "ms_adapter.bin") if not os.path.exists(photomaker_local_path): ms_path = hf_hub_download( @@ -373,50 +375,51 @@ class MSdiffusion_Model_Loader: else: ms_path = photomaker_local_path ms_ckpt = get_instance_path(ms_path) + image_processor = CLIPImageProcessor() - #print(type(image_processor)) - image_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_vision).to(device,dtype=torch.float16) - # print(type(image_encoder)) - image_encoder_projection_dim = image_encoder.config.projection_dim + config_path = os.path.join(dir_path, "config", "config.json") + image_encoder_config = OmegaConf.load(config_path) + image_encoder_projection_dim = image_encoder_config["vision_config"]["projection_dim"] num_tokens = 16 + image_encoder_type = "clip" image_proj_type = "resampler" latent_init_mode = "grounding" # latent_init_mode = "random" + image_proj_model = Resampler( dim=1280, depth=4, dim_head=64, heads=20, num_queries=num_tokens, - embedding_dim=image_encoder.config.hidden_size, + embedding_dim=image_encoder_config["vision_config"]["hidden_size"], # image_encoder.config.hidden_size, output_dim=pipe.unet.config.cross_attention_dim, ff_mult=4, latent_init_mode=latent_init_mode, phrase_embeddings_dim=pipe.text_encoder.config.projection_dim, - ).to(device,dtype=torch.float16) + ).to(device, dtype=torch.float16) ms_model = MSAdapter(pipe.unet, image_proj_model, ckpt_path=ms_ckpt, device=device, num_tokens=num_tokens) ms_model.to(device, dtype=torch.float16) + pipe.to(device, dtype=torch.float16) torch.cuda.empty_cache() - info = ";".join( - [lora, trigger_words,controlnet_model,image_encoder_type,image_proj_type]) - return (pipe,ms_model,image_encoder,image_processor, info,) - + model = {"pipe": pipe, "ms_model": ms_model, "image_encoder": image_encoder, "image_processor": image_processor, + "lora": lora, + "trigger_words": trigger_words, "controlnet_model": controlnet_model, + "image_encoder_type": image_encoder_type, "image_proj_type": image_proj_type, "use_repo": use_repo} + return (model,) + class MSdiffusion_Sampler: def __init__(self): pass - + @classmethod def INPUT_TYPES(cls): return { "required": { "image": ("IMAGE",), - "pipe": ("MODEL",), - "ms_model": ("MODEL",), - "image_encoder": ("MODEL",), - "image_processor": ("MODEL",), - "info": ("STRING", {"forceInput": True, "default": ""}), + "model": ("MSDIF_DICT",), "prompt": ("STRING", {"multiline": True, "default": "a [dog] wearing a pink sunglass"}), "negative_prompt": ("STRING", {"multiline": True, "default": "(worst quality, low quality, normal quality, lowres),"}), @@ -433,9 +436,9 @@ class MSdiffusion_Sampler: "FLOAT", {"default": 0.8, "min": 0.0, "max": 1.0, "step": 0.1, "round": 0.01}), "width": ("INT", {"default": 768, "min": 256, "max": 2048, "step": 64, "display": "number"}), "height": ("INT", {"default": 768, "min": 256, "max": 2048, "step": 64, "display": "number"}), - "batch_size":("INT", {"default": 1, "min": 1, "max": 100,"step": 1,}), - "drop_grounding_tokens":("BOOLEAN", {"default": False},), - "guidance_list": ("STRING", {"multiline": True,"default": "0., 0.25, 0.4, 0.75;0.6, 0.25, 1., 0.75"}), + "batch_size": ("INT", {"default": 1, "min": 1, "max": 100, "step": 1, }), + "drop_grounding_tokens": ("BOOLEAN", {"default": False},), + "guidance_list": ("STRING", {"multiline": True, "default": "0., 0.25, 0.4, 0.75;0.6, 0.25, 1., 0.75"}), }, "optional": {"control_image": ("IMAGE",), @@ -446,17 +449,21 @@ class MSdiffusion_Sampler: def IS_CHANGED(cls, **kwargs): if kwargs.get("batch_size"): return - + RETURN_TYPES = ("IMAGE",) - RETURN_NAMES = ("image", ) + RETURN_NAMES = ("image",) FUNCTION = "ms_sampler" CATEGORY = "MSdiffusion" - - def main_normal(self,prompt,pipe,phrases,ms_model,input_images,num_samples,steps,seed,negative_prompt,scale,image_encoder,cfg,image_processor, - boxes,mask_threshold,start_step,image_proj_type,image_encoder_type,drop_grounding_tokens,height,width,phrase_idxes, eot_idxes): - - images = ms_model.generate(pipe=pipe, pil_images=[input_images], num_samples=num_samples, - num_inference_steps=steps, + + def main_normal(self, prompt, pipe, phrases, ms_model, input_images, num_samples, steps, seed, negative_prompt, + scale, image_encoder, cfg, image_processor, + boxes, mask_threshold, start_step, image_proj_type, image_encoder_type, drop_grounding_tokens, + height, width, phrase_idxes, eot_idxes, image, use_repo): + if use_repo: + image = None + images = ms_model.generate(pipe=pipe, pil_images=[input_images], processed_images=image, + num_samples=num_samples, + num_inference_steps=steps, use_repo=use_repo, seed=seed, prompt=[prompt], negative_prompt=negative_prompt, scale=scale, image_encoder=image_encoder, guidance_scale=cfg, @@ -470,10 +477,16 @@ class MSdiffusion_Sampler: phrase_idxes=phrase_idxes, eot_idxes=eot_idxes, height=height, width=width) return images - def main_control(self,prompt,width,height,pipe,phrases,ms_model,input_images,num_samples,steps,seed,negative_prompt,scale,image_encoder,cfg, - image_processor,boxes,mask_threshold,start_step,image_proj_type,image_encoder_type,drop_grounding_tokens,controlnet_scale,control_image,phrase_idxes, eot_idxes): - images = ms_model.generate(pipe=pipe, pil_images=[input_images], num_samples=num_samples, - num_inference_steps=steps, + + def main_control(self, prompt, width, height, pipe, phrases, ms_model, input_images, num_samples, steps, seed, + negative_prompt, scale, image_encoder, cfg, + image_processor, boxes, mask_threshold, start_step, image_proj_type, image_encoder_type, + drop_grounding_tokens, controlnet_scale, control_image, phrase_idxes, eot_idxes, image, use_repo): + if use_repo: + image = None + images = ms_model.generate(pipe=pipe, pil_images=[input_images], processed_images=image, + num_samples=num_samples, + num_inference_steps=steps, use_repo=use_repo, seed=seed, prompt=[prompt], negative_prompt=negative_prompt, scale=scale, image_encoder=image_encoder, guidance_scale=cfg, @@ -487,10 +500,10 @@ class MSdiffusion_Sampler: phrase_idxes=phrase_idxes, eot_idxes=eot_idxes, height=height, width=width, image=control_image, controlnet_conditioning_scale=controlnet_scale) - + return images - - def get_phrases_idx(self,tokenizer, phrases, prompt): + + def get_phrases_idx(self, tokenizer, phrases, prompt): res = [] phrase_cnt = {} for phrase in phrases: @@ -503,42 +516,53 @@ class MSdiffusion_Sampler: res.append(get_phrase_idx(tokenizer, phrase, prompt, num=cur_cnt)[0]) return res - def get_float(self,str_in): - list_str=str_in.split(",") - float_box=[float(x) for x in list_str] + def get_float(self, str_in): + list_str = str_in.split(",") + float_box = [float(x) for x in list_str] return float_box - - def ms_sampler(self,image,pipe,ms_model,image_encoder,image_processor,info,prompt,negative_prompt,seed, steps, - cfg,scale, mask_threshold, start_step,controlnet_scale,width,height,batch_size,drop_grounding_tokens,guidance_list,**kwargs): - lora, trigger_words,controlnet_model,image_encoder_type,image_proj_type ,= info.split(";") + + def ms_sampler(self, image, model, prompt, negative_prompt, seed, steps, + cfg, scale, mask_threshold, start_step, controlnet_scale, width, height, batch_size, + drop_grounding_tokens, guidance_list, **kwargs): + pipe = model["pipe"] + ms_model = model["ms_model"] + image_encoder = model["image_encoder"] + image_processor = model["image_processor"] + lora = model["lora"] + trigger_words = model["trigger_words"] + controlnet_model = model["controlnet_model"] + image_encoder_type = model["image_encoder_type"] + image_proj_type = model["image_proj_type"] + use_repo = model["use_repo"] if drop_grounding_tokens: - drop_grounding_tokens = [1]# set to 1 if you want to drop the grounding tokens + drop_grounding_tokens = [1] # set to 1 if you want to drop the grounding tokens else: drop_grounding_tokens = [0] - guidance_list=guidance_list.strip().split(";") + guidance_list = guidance_list.strip().split(";") - box_add=[] # 获取预设box + box_add = [] # 获取预设box for i in range(len(guidance_list)): box_add.append(self.get_float(guidance_list[i])) - if mask_threshold==0.: - mask_threshold=None - + if mask_threshold == 0.: + mask_threshold = None + if '[' in prompt and ']' in prompt: object_prompt = extract_content_from_brackets(prompt) # 提取prompt的object list - unique_object_prompt = sorted(list(set(object_prompt)),key=lambda x: list(object_prompt).index(x)) # 清除同名物体,保持原有顺序 + unique_object_prompt = sorted(list(set(object_prompt)), + key=lambda x: list(object_prompt).index(x)) # 清除同名物体,保持原有顺序 object_num = len(unique_object_prompt) else: raise "Need to enclose the name of the object in square brackets" d1, _, _, _ = image.size() - #print(d1,object_num) + # print(d1,object_num) if lora != "none": prompt = prompt + trigger_words + " " + "style" - prompt=prompt.replace("]", " ").replace("[", " ") #去除[] + prompt = prompt.replace("]", " ").replace("[", " ") # 去除[] if object_num == d1: phrases = [unique_object_prompt] @@ -549,7 +573,7 @@ class MSdiffusion_Sampler: if d1 == 1: input_images = [nomarl_upscale_topil(image, 512, 512)] torch.cuda.empty_cache() - if mask_threshold: # step之前所用布局指导 + if mask_threshold: # step之前所用布局指导 boxes = box_add[0] boxes = [[boxes]] # 随机选1个自定义的 # boxes = [[[0.25, 0.25, 0.75, 0.75]]] # 1 object @@ -559,57 +583,66 @@ class MSdiffusion_Sampler: eot_idxes = [[get_eot_idx(pipe.tokenizer, prompt)] * len(phrases[0])] if controlnet_model != "none": control_image = kwargs.get("control_image") - control_img = nomarl_upscale_topil(control_image, width, height) - image_main=self.main_control(prompt,width,height,pipe,phrases,ms_model,input_images,batch_size,steps,seed,negative_prompt,scale,image_encoder,cfg, - image_processor,boxes,mask_threshold,start_step,image_proj_type,image_encoder_type,drop_grounding_tokens,controlnet_scale,control_img,phrase_idxes, eot_idxes) - - else: #no controlnet - image_main=self.main_normal( prompt, pipe, phrases, ms_model, input_images, batch_size, steps, seed, - negative_prompt, scale, image_encoder, cfg, image_processor, - boxes, mask_threshold, start_step, image_proj_type, image_encoder_type, - drop_grounding_tokens, height, width,phrase_idxes, eot_idxes) - + control_img = nomarl_upscale_topil(control_image, width, height) + image_main = self.main_control(prompt, width, height, pipe, phrases, ms_model, input_images, batch_size, + steps, seed, negative_prompt, scale, image_encoder, cfg, + image_processor, boxes, mask_threshold, start_step, image_proj_type, + image_encoder_type, drop_grounding_tokens, controlnet_scale, control_img, + phrase_idxes, eot_idxes, image, use_repo) + + else: # no controlnet + image_main = self.main_normal(prompt, pipe, phrases, ms_model, input_images, batch_size, steps, seed, + negative_prompt, scale, image_encoder, cfg, image_processor, + boxes, mask_threshold, start_step, image_proj_type, image_encoder_type, + drop_grounding_tokens, height, width, phrase_idxes, eot_idxes, image, + use_repo) + else: img_list = list(torch.chunk(image, chunks=d1)) - #print(img_list) + # print(img_list) input_images = [nomarl_upscale_topil(img, 512, 512) for img in img_list] - #print(input_images) + # print(input_images) if mask_threshold: - if len(box_add)