# import os # render2real_path = "/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/spotlight_sketch/epoch0" # sketch_enhance_body_path = "/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/spotlight_sketch/GT" # render2real_files = set(os.listdir(render2real_path)) # sketch_enhance_body_files = set(os.listdir(sketch_enhance_body_path)) # for file in render2real_files: # if file not in sketch_enhance_body_files: # print(f"Removing {file} from render2real_path") # os.remove(os.path.join(render2real_path, file)) # for file in sketch_enhance_body_files: # if file not in render2real_files: # print(f"Removing {file} from sketch_enhance_body_path") # os.remove(os.path.join(sketch_enhance_body_path, file)) # import os # input_path = "/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/mbti/Realistic" # for file in os.listdir(input_path): # # 把文件名称最后的_Realistic去掉 # new_name = file.replace("_Realistic", "") # os.rename(os.path.join(input_path, file), os.path.join(input_path, new_name)) # import os # for file in os.listdir("dataset/spotlight_sketch_cat/GT"): # with open("dataset/spotlight_sketch_cat/pairs_t2t.txt", "a") as f: # # 目标图 原图 # f.write(f"GT/{file}\tepoch0/{file}\n") # import os # import json # from tqdm import tqdm # input_txt = "/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/spotlight_sketch_cat/spotlight_nano_comprehension_1203.txt" # with open(input_txt, "r") as f: # lines = f.readlines() # for i in tqdm(range(len(lines))): # data = json.loads(lines[i]) # fig_id = f"{data['Image_Name']}.png" # del data["Image_Name"] # input_dir = "dataset/spotlight_sketch_cat/epoch0" # GT_dir = "dataset/spotlight_sketch_cat/GT" # for file in os.listdir(input_dir): # if fig_id in file: # with open("dataset/spotlight_sketch_cat/pairs_i2i.txt", "a") as f: # # 目标 原图 prompt # f.write(f"{GT_dir}/{file}\t{input_dir}/{file}\t{data}\n") # 把文件夹中的图片每六张拼成一个3行两列的大图，保存到另一个文件夹中，原图拼接不要截图 import os from PIL import Image from tqdm import tqdm import numpy as np base_dirs = ["/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/the roses","/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/nouvelle","/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/legs","/fi-lib/workspace/sjx/DiffSynth-Studio/dataset/frankenstein"] # 核心配置参数（按需求定义） crop_size = (1920, 800) # 目标CenterCrop尺寸 (宽, 高) resize_size = (477, 188) # 下采样后的单张尺寸 (宽, 高) line_width = 6 # 黑线宽度（6像素） target_merge_size = (960, 576) # 最终拼接目标尺寸 (宽, 高) def center_crop_to_size(img, target_size): """ 对图片进行CenterCrop到指定尺寸，不足部分用黑色像素填充 :param img: PIL Image对象 :param target_size: (target_w, target_h) 目标裁剪尺寸 :return: crop+补黑后的PIL Image """ target_w, target_h = target_size img_w, img_h = img.size # Step1: 计算CenterCrop的区域（中心对齐） # 水平方向裁剪 if img_w >= target_w: left = (img_w - target_w) // 2 right = left + target_w else: left = 0 right = img_w # 垂直方向裁剪 if img_h >= target_h: top = (img_h - target_h) // 2 bottom = top + target_h else: top = 0 bottom = img_h # Step2: 执行CenterCrop cropped = img.crop((left, top, right, bottom)) # Step3: 不足目标尺寸的部分用黑色填充 if cropped.size != (target_w, target_h): new_img = Image.new("RGB", (target_w, target_h), (0, 0, 0)) # 黑色背景 new_img.paste(cropped, ((target_w - cropped.width) // 2, (target_h - cropped.height) // 2)) cropped = new_img return cropped for k in range(len(base_dirs)): save_path = f"{base_dirs[k]}_dedup_cat" os.makedirs(save_path, exist_ok=True) input_path = f"{base_dirs[k]}_dedup" # 获取并排序文件列表 files = [f for f in os.listdir(input_path) if f.lower().endswith(('.png', '.jpg', '.jpeg'))] files.sort() # 遍历文件，每6张拼接一次 for i in tqdm(range(0, len(files), 6), desc="拼接图片"): # 初始化最终拼接画布（目标尺寸960×576，黑色背景） merged_image = np.zeros((target_merge_size[1], target_merge_size[0], 3), dtype=np.uint8) # 逐个处理6张图片 valid_imgs = [] # 存储处理后的有效图片 for j in range(6): if i + j >= len(files): # 不足6张时，break img_np = np.zeros((resize_size[1], resize_size[0], 3), dtype=np.uint8) valid_imgs.append(img_np) continue img_path = os.path.join(input_path, files[i + j]) try: # 读取图片（确保RGB格式） img = Image.open(img_path).convert("RGB") img_w, img_h = img.size # 过滤条件：原始宽度不足1800则跳过并打印 if img_w < 1800: print(f"跳过文件 {files[i+j]}: 原始宽度 {img_w} < 1800") # 用黑色图片填充该位置 img_np = np.zeros((resize_size[1], resize_size[0], 3), dtype=np.uint8) valid_imgs.append(img_np) continue # Step1: CenterCrop到1920×800，不足补黑 cropped_img = center_crop_to_size(img, crop_size) # Step2: 下采样到477×188（LANCZOS插值，保持画质） resized_img = cropped_img.resize(resize_size, resample=Image.LANCZOS) # 转为numpy数组 img_np = np.array(resized_img) valid_imgs.append(img_np) except Exception as e: print(f"处理文件 {files[i+j]} 出错: {str(e)}") # 出错时用黑色图片填充 img_np = np.zeros((resize_size[1], resize_size[0], 3), dtype=np.uint8) valid_imgs.append(img_np) # Step3: 计算每张图在拼接画布中的位置（3行2列 + 6像素黑线） # 验证拼接尺寸兼容性（防止配置错误） assert len(valid_imgs) == 6, "有效图片数量必须为6张" # 计算图片+黑线的总占位，确保适配960×576 total_col = 2 * resize_size[0] + 1 * line_width # 2列+1条竖线 total_row = 3 * resize_size[1] + 2 * line_width # 3行+2条横线 # 计算画布中的偏移（居中放置，保证最终尺寸960×576） offset_x = (target_merge_size[0] - total_col) // 2 offset_y = (target_merge_size[1] - total_row) // 2 # 逐个放置图片到拼接画布 for idx, img_np in enumerate(valid_imgs): row = idx // 2 # 0/1/2行 col = idx % 2 # 0/1列 # 计算当前图片的起始位置（含黑线+整体偏移） x_start = offset_x + col * (resize_size[0] + line_width) y_start = offset_y + row * (resize_size[1] + line_width) x_end = x_start + resize_size[0] y_end = y_start + resize_size[1] # 确保不超出画布边界 x_end = min(x_end, target_merge_size[0]) y_end = min(y_end, target_merge_size[1]) x_start = max(x_start, 0) y_start = max(y_start, 0) # 放置图片到画布 merged_image[y_start:y_end, x_start:x_end, :] = img_np[:y_end-y_start, :x_end-x_start, :] # Step4: 保存最终拼接图片 save_name = f'merged_{i//6}.png' save_full_path = os.path.join(save_path, save_name) Image.fromarray(merged_image).save(save_full_path) print(f"所有图片处理完成！结果保存至: {save_path}")