""" XB-BOX AI 漫画文字渲染节点 (生产版 v3.0) ========================================= 基于"文本光栅化物理渲染 + CV视觉锚定 + 前端提示词分流"三层架构， 从根本上解决扩散模型生成中文对白时的笔画畸形、鬼影残留与排版错位问题。 节点清单: XB_ComicPromptParser — 前端：智能提示词解析分流 + 字数-气泡尺寸映射 XB_ComicTextRenderer — 第一代：手动精确坐标渲染 (旁白/拟声词降级预案) XB_AutoBubbleTextRenderer — 第二代：全自动 CV 寻迹 + 数字涂改液 + 三层防御 """ import os import re import textwrap import numpy as np import torch from PIL import Image, ImageDraw, ImageFont try: import cv2 HAS_CV2 = True except ImportError: HAS_CV2 = False # ============================================================ # 字体工具 # ============================================================ def _find_cjk_fonts(): """扫描系统常见中文字体，返回可用字体路径列表""" font_paths = [] win_font_dir = os.environ.get("WINDIR", "C:\\Windows") + "\\Fonts" for fname in ["msyh.ttc", "msyhbd.ttc", "simhei.ttf", "simsun.ttc", "simkai.ttf", "Deng.ttf", "simfang.ttf"]: fp = os.path.join(win_font_dir, fname) if os.path.exists(fp): font_paths.append(fp) # Linux / macOS fallback if not font_paths: for d in ["/usr/share/fonts/truetype/", "/System/Library/Fonts/"]: if os.path.isdir(d): for f in os.listdir(d): if f.endswith((".ttf", ".ttc", ".otf")): font_paths.append(os.path.join(d, f)) return font_paths def _load_font(font_name, font_size): """加载字体：支持直接路径、文件名模糊搜索、系统回退""" if not font_name: font_name = "" if os.path.isfile(font_name): try: return ImageFont.truetype(font_name, font_size) except Exception: pass for fp in _find_cjk_fonts(): if font_name.lower() in os.path.basename(fp).lower(): try: return ImageFont.truetype(fp, font_size) except Exception: continue for fp in _find_cjk_fonts(): try: return ImageFont.truetype(fp, font_size) except Exception: continue return ImageFont.load_default() def _parse_hex(color_str): """#RRGGBB → (R, G, B, 255)""" c = color_str.lstrip("#") if len(c) >= 6: return (int(c[0:2], 16), int(c[2:4], 16), int(c[4:6], 16), 255) return (0, 0, 0, 255) def _wrap_text(text, font, max_width, draw): """按最大宽度自动换行，返回行列表""" if not text or max_width <= 0: return [text] if text else [] lines = [] for para in text.split("\n"): if not para: lines.append("") continue cur = "" for ch in para: test = cur + ch bw = draw.textbbox((0, 0), test, font=font)[2] if bw > max_width and cur: lines.append(cur) cur = ch else: cur = test if cur: lines.append(cur) return lines # ============================================================ # 节点 A: XB_ComicPromptParser — 前端提示词智能分流 # ============================================================ class XB_ComicPromptParser: """从含 [台词] 标记的剧本中提取台词并清洗提示词。 自动根据字数向大模型注入匹配的气泡尺寸描述，实现 SSOT 单一数据源架构。""" @classmethod def INPUT_TYPES(cls): return { "required": { "raw_prompt": ("STRING", { "default": "第一格大图，男生冷酷，[宋琛眉眼清秀，嘴里却叼着根突兀烟]，左侧气泡，[真难看。]", "multiline": True, "tooltip": "完整剧本。用方括号 [台词] 标记对白，节点会自动提取和清洗" }), "base_bubble_desc": ("STRING", { "default": "纯白干净的空白对话框", "tooltip": "气泡基础描述。节点会根据字数自动追加微小/标准/宽大/超大等修饰词" }), }, } RETURN_TYPES = ("STRING", "STRING") RETURN_NAMES = ("洗净给大模型的提示词", "给渲染节点的台词(|分隔)") FUNCTION = "parse" CATEGORY = "XB_ToolBox/Comic" DESCRIPTION = ( "前端提示词拦截解析器 —— SSOT 架构核心。\n" "输入含 [台词] 的完整剧本，自动:\n" "1. 提取 [] 内台词，拼接为台词1|台词2 格式\n" "2. 根据字数映射气泡尺寸 (<=3:微小, <=8:标准, <=18:宽大, >18:超大)\n" "3. 清洗提示词，彻底屏蔽大模型看到汉字，根除鬼影" ) def parse(self, raw_prompt, base_bubble_desc): dialogues = [] def replacer(m): text = m.group(1).strip() if not text: return "" dialogues.append(text) n = len(text) if n <= 3: adj = "一个微小紧凑的" elif n <= 8: adj = "一个标准大小的" elif n <= 18: adj = "一个宽大的" else: adj = "一个占据极大面积的超大长条形" return adj + base_bubble_desc cleaned = re.sub(r"\[(.*?)\]", replacer, raw_prompt, flags=re.DOTALL) pipeline = " | ".join(dialogues) print(f"[XB_ComicPromptParser] 提取 {len(dialogues)} 句台词，已按字数映射气泡尺寸") return (cleaned, pipeline) # ============================================================ # 节点 B: XB_ComicTextRenderer — 手动精确坐标 (降级预案) # ============================================================ class XB_ComicTextRenderer: """在图像指定坐标用系统字体精确渲染文字。 用于全自动节点无法覆盖的场景：旁白、拟声词、无气泡文本。""" @classmethod def INPUT_TYPES(cls): return { "required": { "image": ("IMAGE",), "text": ("STRING", {"default": "", "multiline": True, "tooltip": "要渲染的文字。支持 \\n 换行"}), "font_name": ("STRING", {"default": "msyh.ttc", "tooltip": "字体名或路径，如 msyh.ttc / simhei.ttf"}), "font_size": ("INT", {"default": 36, "min": 8, "max": 500, "step": 1}), "x": ("INT", {"default": 100, "min": 0, "max": 8192, "step": 1, "tooltip": "文字起始 X 坐标"}), "y": ("INT", {"default": 100, "min": 0, "max": 8192, "step": 1, "tooltip": "文字起始 Y 坐标"}), "max_width": ("INT", {"default": 400, "min": 0, "max": 8192, "step": 1, "tooltip": "最大宽度(px)，超宽自动换行；0=不限制"}), "alignment": (["left", "center", "right"], {"default": "center"}), "text_color": ("STRING", {"default": "#000000"}), "outline_width": ("INT", {"default": 0, "min": 0, "max": 10, "step": 1}), "outline_color": ("STRING", {"default": "#FFFFFF"}), }, } RETURN_TYPES = ("IMAGE",) RETURN_NAMES = ("图像",) FUNCTION = "render" CATEGORY = "XB_ToolBox/Comic" DESCRIPTION = ( "第一代物理光栅化节点 —— 手动指定 X/Y 坐标，用系统原生字体精确渲染。\n" "用于处理旁白、拟声词、无气泡文本等全自动节点无法覆盖的特殊场景。\n" "提示: 如需多句不同位置，请串联多个此节点分层叠加。" ) def render(self, image, text, font_name, font_size, x, y, max_width, alignment, text_color, outline_width, outline_color): fg = _parse_hex(text_color) og = _parse_hex(outline_color) font = _load_font(font_name, font_size) B = image.shape[0] out = [] for b in range(B): arr = (image[b].cpu().numpy() * 255).astype(np.uint8) pil = Image.fromarray(arr).convert("RGBA") draw = ImageDraw.Draw(pil) lines = _wrap_text(text, font, max_width, draw) bb = draw.textbbox((0, 0), "啊", font=font) lh = bb[3] - bb[1] + 4 for i, line in enumerate(lines): lw = draw.textbbox((0, 0), line, font=font)[2] if alignment == "center" and max_width > 0: lx = x + (max_width - lw) // 2 elif alignment == "right" and max_width > 0: lx = x + max_width - lw else: lx = x ly = y + i * lh if outline_width > 0: for dx in range(-outline_width, outline_width + 1): for dy in range(-outline_width, outline_width + 1): if dx or dy: draw.text((lx + dx, ly + dy), line, font=font, fill=og) draw.text((lx, ly), line, font=font, fill=fg) result = np.array(pil.convert("RGB")).astype(np.float32) / 255.0 out.append(torch.from_numpy(result)) return (torch.stack(out, dim=0),) # ============================================================ # 节点 C: XB_AutoBubbleTextRenderer — 全自动 CV 寻迹 + 涂改液 # ============================================================ class XB_AutoBubbleTextRenderer: """OpenCV 自动扫描白色对话框 -> 涂改液抹除鬼影 -> 填入台词。 含三层防御: 暗黑模式 / 形状过滤 / 边界分离。 形态学预处理解决气泡内纹理导致的识别断裂。""" @classmethod def INPUT_TYPES(cls): return { "required": { "image": ("IMAGE",), "text": ("STRING", {"default": "", "multiline": True, "tooltip": "台词，竖线 | 分隔。如: 台词1 | 台词2"}), "font_name": ("STRING", {"default": "msyh.ttc", "tooltip": "字体名或路径，如 msyh.ttc / simhei.ttf"}), "font_size": ("INT", {"default": 36, "min": 8, "max": 500, "step": 1}), "text_color": ("STRING", {"default": "#000000"}), "bubble_margin": ("INT", {"default": 15, "min": 0, "max": 200, "step": 1}), "auto_clear_bubble": ("BOOLEAN", { "default": True, "tooltip": "数字涂改液: 用纯白填满气泡内部，彻底抹除 AI 鬼影乱码后再印字" }), "white_threshold": ("INT", {"default": 220, "min": 160, "max": 255, "step": 1, "tooltip": "白色阈值。降低可检测带光影的气泡"}), "min_bubble_area": ("INT", {"default": 5000, "min": 500, "max": 1000000, "step": 100}), "sort_mode": (["auto", "top_to_bottom", "left_to_right", "largest_first"], {"default": "auto"}), "invert_mode": ("BOOLEAN", {"default": False, "tooltip": "暗黑模式: 取反后检测深色气泡"}), "shape_filter_enabled": ("BOOLEAN", {"default": True, "tooltip": "形状过滤: 排除白T恤等长条形误检"}), "min_extent": ("FLOAT", {"default": 0.4, "min": 0.1, "max": 1.0, "step": 0.05}), "max_aspect_ratio": ("FLOAT", {"default": 3.5, "min": 1.0, "max": 10.0, "step": 0.1}), "erode_iterations": ("INT", {"default": 1, "min": 0, "max": 5, "step": 1, "tooltip": "边界分离: 腐蚀迭代，分离粘连气泡"}), "morph_close_size": ("INT", {"default": 15, "min": 3, "max": 51, "step": 2, "tooltip": "形态学内核尺寸。增大可填平气泡内的噪点和残影"}), }, } RETURN_TYPES = ("IMAGE", "STRING") RETURN_NAMES = ("图像", "检测信息") FUNCTION = "render_auto" CATEGORY = "XB_ToolBox/Comic" DESCRIPTION = ( "第二代全自动 CV 视觉锚定节点 (v3.0 生产版)。\n\n" "流程: 灰度化 -> 形态学闭运算填平噪点 -> 二值化 -> 轮廓检测 -> 涂改液抹白 -> 填字。\n\n" "核心能力:\n" "- auto_clear_bubble: 数字涂改液，先抹后写彻底根除鬼影\n" "- morph_close_size: 形态学预处理，解决气泡内纹理/残影导致的识别断裂\n" "- 三层防御: 暗黑模式(invert) + 形状过滤(shape_filter) + 边界分离(erode)\n" "- 台词框输入 台词1 | 台词2 | 台词3，一键自动排版" ) def render_auto(self, image, text, font_name, font_size, text_color, bubble_margin, auto_clear_bubble, white_threshold, min_bubble_area, sort_mode, invert_mode=False, shape_filter_enabled=True, min_extent=0.4, max_aspect_ratio=3.5, erode_iterations=1, morph_close_size=15): if not HAS_CV2: raise ImportError( "[XB_AutoBubbleTextRenderer] 需要 opencv-python。请运行: pip install opencv-python") fg = _parse_hex(text_color) dialogues = [d.strip() for d in text.split("|") if d.strip()] font = _load_font(font_name, font_size) B = image.shape[0] results = [] info = [] for b in range(B): # ---- torch -> numpy -> BGR ---- arr = (image[b].cpu().numpy() * 255).astype(np.uint8) if arr.shape[2] == 4: cv_rgb = cv2.cvtColor(arr, cv2.COLOR_RGBA2RGB) else: cv_rgb = arr.copy() cv_bgr = cv2.cvtColor(cv_rgb, cv2.COLOR_RGB2BGR) # ---- 1. 灰度 + 暗黑模式 ---- gray = cv2.cvtColor(cv_bgr, cv2.COLOR_BGR2GRAY) if invert_mode: gray = cv2.bitwise_not(gray) info.append(" 🌙 暗黑模式: 已取反") # ---- 2. 形态学预处理 (关键!) ---- # 先用大核闭运算把气泡内 AI 残影/噪点/纹理糊成一片 mk = cv2.getStructuringElement(cv2.MORPH_RECT, (morph_close_size, morph_close_size)) gray = cv2.morphologyEx(gray, cv2.MORPH_CLOSE, mk) # ---- 3. 二值化 ---- _, binary = cv2.threshold(gray, white_threshold, 255, cv2.THRESH_BINARY) # ---- 4. 边界分离 (开运算) ---- if erode_iterations > 0: ek = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7)) binary = cv2.morphologyEx(binary, cv2.MORPH_OPEN, ek, iterations=erode_iterations) # ---- 5. 二次闭运算 (弥合腐蚀造成的小缺口) ---- ck = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5)) binary = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, ck, iterations=2) # ---- 6. 轮廓检测 ---- contours, _ = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) # ---- 7. 筛选 (面积 + 形状先验) ---- filtered = 0 bubbles = [] for cnt in contours: area = cv2.contourArea(cnt) if area < min_bubble_area: continue rx, ry, rw, rh = cv2.boundingRect(cnt) if shape_filter_enabled: bbox_area = float(rw * rh) extent = area / bbox_area if bbox_area > 0 else 0.0 if extent < min_extent: filtered += 1 continue ar = max(rw, rh) / min(rw, rh) if min(rw, rh) > 0 else 999.0 if ar > max_aspect_ratio: filtered += 1 continue bubbles.append({ "x": rx, "y": ry, "width": rw, "height": rh, "cx": rx + rw // 2, "cy": ry + rh // 2, "area": area, "contour": cnt, }) if filtered: info.append(f" 🛡️ 形状过滤: 排除 {filtered} 个误检") # ---- 8. 排序 ---- if sort_mode == "top_to_bottom": bubbles.sort(key=lambda b: (b["cy"], b["cx"])) elif sort_mode == "left_to_right": bubbles.sort(key=lambda b: (b["cx"], b["cy"])) elif sort_mode == "largest_first": bubbles.sort(key=lambda b: -b["area"]) else: if bubbles: avg_h = sum(b["height"] for b in bubbles) / len(bubbles) row_th = avg_h * 0.5 bubbles.sort(key=lambda b: b["cx"]) rows = [] for bu in bubbles: placed = False for row in rows: if abs(bu["cy"] - row[0]["cy"]) < row_th: row.append(bu) placed = True break if not placed: rows.append([bu]) for row in rows: row.sort(key=lambda b: b["cx"]) rows.sort(key=lambda r: r[0]["cy"]) bubbles = [b for row in rows for b in row] info.append(f"🫧 Batch {b+1}: 检测 {len(bubbles)} 气泡, 台词 {len(dialogues)} 句") if not bubbles: info.append(" ⚠️ 未检测到气泡! 请降低 white_threshold 或增大 morph_close_size") results.append(torch.from_numpy(arr.astype(np.float32) / 255.0)) continue # ---- 9. 涂改液: 用轮廓填充纯白 ---- if auto_clear_bubble: valid_cnts = [b["contour"] for b in bubbles] cv2.drawContours(cv_rgb, valid_cnts, -1, (255, 255, 255), thickness=cv2.FILLED) info.append(f" 🧼 涂改液: 擦除 {len(valid_cnts)} 个气泡内部鬼影") # ---- 10. PIL 渲染文字 ---- pil = Image.fromarray(cv_rgb) draw = ImageDraw.Draw(pil) for i, bubble in enumerate(bubbles): if i >= len(dialogues): break line_text = dialogues[i] avail_w = max(font_size, bubble["width"] - 2 * bubble_margin) text_lines = _wrap_text(line_text, font, avail_w, draw) bb = draw.textbbox((0, 0), "啊", font=font) lh = bb[3] - bb[1] + 4 total_h = len(text_lines) * lh start_y = bubble["y"] + (bubble["height"] - total_h) // 2 for j, tline in enumerate(text_lines): tlw = draw.textbbox((0, 0), tline, font=font)[2] tlx = bubble["x"] + (bubble["width"] - tlw) // 2 tly = start_y + j * lh draw.text((tlx, tly), tline, font=font, fill=fg) info.append( f" 💬 气泡{i+1}: ({bubble['x']},{bubble['y']}) " f"{bubble['width']}x{bubble['height']} -> \"{line_text[:20]}{'...' if len(line_text) > 20 else ''}\"" ) result = np.array(pil).astype(np.float32) / 255.0 results.append(torch.from_numpy(result)) out_info = "\n".join(info) print(f"[XB_AutoBubbleTextRenderer]\n{out_info}") return (torch.stack(results, dim=0), out_info)