#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ AI 漫剧 · 剧本导入器 ===================== 把小说/剧本文本： 1) 按「第X章/回/节」拆分章节 2) 每章按目标字数拆成若干「集」 3) 用 LLM 提取人物/场景/道具等资产（无 key 时退化为规则提取） 4) 生成可导入画布的 JSON（{cards, view}） 用法: python3 aicomic-import.py 小说.txt # 默认输出 aicomic-import.json python3 aicomic-import.py 小说.txt -o out.json # 指定输出 python3 aicomic-import.py 小说.txt --target 1500 # 每集目标字数 python3 aicomic-import.py 小说.txt --no-llm # 只用规则，不调 AI """ import json import os import re import sys BASE_DIR = os.path.dirname(os.path.abspath(__file__)) CONFIG_PATH = os.path.join(BASE_DIR, 'manju', 'api_config.json') # ---------------- 1) 拆章 ---------------- HEAD_RE = re.compile( r'^\s*(第\s*[0-9一二三四五六七八九十百千零两]+\s*[章回节卷]' r'|楔子|序章|序言|引子|尾声|番外|终章|后记|大结局' r'|(?:CHAPTER|Chapter)\s+[0-9IVXLC]+)' r'[\s:：、.．]*([^\n]{0,30})?' ) def split_chapters(text): """按章节标题拆章。无标题时整篇算一章。返回 [{title, content}]""" lines = text.split('\n') chapters = [] cur_title = None cur = [] for ln in lines: ln = re.sub(r'^#+\s*', '', ln) # 兼容 markdown 标题前缀 m = HEAD_RE.match(ln.strip()) if m: if cur_title is not None: chapters.append({'title': cur_title, 'content': '\n'.join(cur).strip()}) cur_title = (m.group(1).strip() + (' ' + m.group(2).strip() if m.group(2) else '')).strip() cur = [] else: cur.append(ln) if cur_title is not None: chapters.append({'title': cur_title, 'content': '\n'.join(cur).strip()}) if not chapters: chapters = [{'title': '正文', 'content': text.strip()}] chapters = [c for c in chapters if c['content'].strip()] return chapters # ---------------- 2) 拆集 ---------------- def split_episodes(text, target_chars=1500): """按段落边界把一章拆成若干集，每集目标 target_chars 字""" paras = [p.strip() for p in re.split(r'\n\s*\n', text) if p.strip()] if not paras: paras = [p.strip() for p in text.splitlines() if p.strip()] episodes, cur, curlen = [], [], 0 for p in paras: if cur and curlen >= target_chars: episodes.append('\n'.join(cur)) cur, curlen = [], 0 cur.append(p) curlen += len(p) if cur: episodes.append('\n'.join(cur)) return episodes or ([text.strip()] if text.strip() else []) # ---------------- 3) LLM 提取资产 ---------------- def _load_llm_config(): cfg = {} if os.path.exists(CONFIG_PATH): try: cfg = json.load(open(CONFIG_PATH, encoding='utf-8')) except Exception: cfg = {} providers = [] if cfg.get('api_key'): providers.append({ 'name': 'agnes', 'base': cfg.get('base_url', 'https://apihub.agnes-ai.com/v1'), 'key': cfg['api_key'], 'model': cfg.get('llm_model', 'agnes-2.5-flash')}) if os.environ.get('DEEPSEEK_API_KEY'): providers.append({ 'name': 'deepseek', 'base': 'https://api.deepseek.com/v1', 'key': os.environ['DEEPSEEK_API_KEY'], 'model': 'deepseek-chat'}) cred = os.path.expanduser('~/.dsh/.credentials.yaml') if os.path.exists(cred): try: for ln in open(cred, encoding='utf-8'): if ln.startswith('DEEPSEEK_API_KEY'): k = ln.split(':', 1)[1].strip().strip(chr(34)).strip(chr(39)) if k: providers.append({ 'name': 'deepseek', 'base': 'https://api.deepseek.com/v1', 'key': k, 'model': 'deepseek-chat'}) except Exception: pass return providers def _llm_chat(providers, system, user, max_tokens=2500): import urllib.request errs = [] for p in providers: body = json.dumps({ 'model': p['model'], 'messages': [{'role': 'system', 'content': system}, {'role': 'user', 'content': user}], 'max_tokens': max_tokens, 'temperature': 0.3}).encode('utf-8') req = urllib.request.Request(p['base'].rstrip('/') + '/chat/completions', data=body, headers={ 'Content-Type': 'application/json', 'Authorization': 'Bearer ' + p['key']}) try: with urllib.request.urlopen(req, timeout=120) as r: d = json.loads(r.read().decode('utf-8')) return d['choices'][0]['message']['content'] except Exception as e: errs.append('%s: %s' % (p['name'], str(e)[:150])) raise RuntimeError('LLM 全部不可用: ' + ' | '.join(errs)) def _extract_json(text): """从 LLM 回复里宽容地抠出 JSON 对象""" text = text.strip() if text.startswith(chr(96) * 3): text = re.sub('^' + chr(96) * 3 + '[a-zA-Z]*\n', '', text) text = re.sub('\n' + chr(96) * 3 + '$', '', text) try: return json.loads(text) except Exception: pass s, e = text.find('{'), text.rfind('}') if s >= 0 and e > s: try: return json.loads(text[s:e + 1]) except Exception: pass return None def extract_assets_llm(chapters, providers): """LLM 提取资产，返回 {'title','characters':[...],'scenes':[...],'props':[...]} 通读每章开头较多内容（合计最多约10万字），保证不遗漏。""" parts = [] for c in chapters: take = min(len(c['content']), 40000) parts.append(c['title'] + '\n' + c['content'][:take]) context = '\n\n'.join(parts)[:100000] system = ('你是资深网文编辑与漫剧制作人。请通读下面小说节选，' '完整提取制作漫剧所需的结构化信息。只输出一个 JSON 对象，不要任何其他文字。') user = ('小说节选（较长，请通读全部内容，不要只看开头）：\n' + context + '\n\n' '请输出 JSON，字段如下：\n' '{\n' ' "title": "小说/剧本名",\n' ' "characters": [{"name": "角色名", "role": "主角/重要配角/反派/路人等", "description": "外貌与性格，20字内", "outfit": "标志性服装，如：灰色卫衣+牛仔裤", "feature": "标志特征，如：左眼角泪痣"}],\n' ' "scenes": ["场景名，如：林府大门·清晨"],\n' ' "props": ["关键道具名，如：木剑"]\n' '}\n' '要求：\n' '1. 完整提取：所有出场过且有名字/有戏份的角色都要列入，配角、反派、关键路人也不要遗漏；\n' '2. 所有出现过的主要地点都要列入场景；\n' '3. 反复出现或推动剧情的关键道具都要列入；\n' '4. 上限：characters 最多25个，scenes 最多25个，props 最多20个，宁多勿漏；\n' '5. 角色名必须从文中出现的人名提取，不要编造。') raw = _llm_chat(providers, system, user) d = _extract_json(raw) or {} d.setdefault('characters', []) d.setdefault('scenes', []) d.setdefault('props', []) d.setdefault('title', '未命名剧本') if isinstance(d.get('characters'), list): fixed = [] for c in d['characters']: if isinstance(c, dict) and c.get('name'): fixed.append({'name': str(c['name']).strip(), 'role': str(c.get('role', '')).strip(), 'description': str(c.get('description', '')).strip(), 'outfit': str(c.get('outfit', '')).strip(), 'feature': str(c.get('feature', '')).strip()}) elif isinstance(c, str) and c.strip(): fixed.append({'name': c.strip(), 'role': '', 'description': '', 'outfit': '', 'feature': ''}) d['characters'] = fixed return d STYLE_PACKS = { 'cel': {'name': '日系赛璐璐（默认）', 'en': 'japanese anime cel-shaded key visual, clean bold lineart, vibrant flat colors, ' 'cinematic rim lighting, soft shadows, expressive anime eyes, professional anime illustration'}, 'guoman': {'name': '国漫厚涂', 'en': 'chinese manhua thick-paint style, rich painterly brushwork, dramatic chiaroscuro, ' 'vivid saturated colors, expressive character art, professional manhua illustration'}, 'ink': {'name': '水墨仙侠', 'en': 'chinese ink-wash painting style, ethereal mist, flowing sumi-e brushstrokes, ' 'xianxia fantasy atmosphere, poetic negative space, high detail'}, 'realistic': {'name': '写实电影感', 'en': 'cinematic photorealistic style, dramatic volumetric lighting, film grain, ' 'shallow depth of field, 8k, hyper-detailed, professional film still'}, '3d': {'name': '3D渲染', 'en': 'high quality 3d render, soft global illumination, subsurface scattering, ' 'detailed character model, octane render, cinematic'}, } QUALITY_TOKENS = { 'std': 'high quality, detailed, sharp focus', 'high': 'masterpiece, best quality, ultra-detailed, intricate details, 8k, ' 'sharp focus, professional illustration, award winning', } def optimize_image_prompt(scene='', detail='', chars=None, move='', style='', style_pack='cel', quality='high'): """把中文分镜改写成英文画面描述（LLM），并由系统统一追加精修画风包+质量词， 保证多镜头画风稳定、人物细节出色。""" chars = chars or [] scene = (scene or '').strip()[:300] pack = STYLE_PACKS.get(style_pack) or STYLE_PACKS['cel'] style_en = pack['en'] quality_en = QUALITY_TOKENS.get(quality, QUALITY_TOKENS['high']) # 中文角色信息（喂给 LLM 翻译进画面描述） char_info = [] for c in chars: nm = c.get('name', '') parts = [] if c.get('appearance'): parts.append('外貌：' + c['appearance']) if c.get('outfit'): parts.append('服装：' + c['outfit']) if c.get('feature'): parts.append('标志特征：' + c['feature']) if nm and parts: char_info.append('角色「%s」：%s' % (nm, '，'.join(parts))) elif nm: char_info.append('角色「%s」' % nm) visual = '' providers = _load_llm_config() if providers: user = ('请把下面的漫剧分镜信息改写成一段用于 AI 生图的英文画面描述（2~4句，只写画面本身：' '人物外貌与动作、景别（近景/中景/全景）、光线、氛围、构图）。' '出场角色的外貌/服装/标志特征必须准确写进描述中，同一角色每次描述保持一致。' '不要加画风词、不要加质量词、不要加"vertical composition"等词，这些由系统统一追加。' '直接输出英文，不要任何解释。\n\n' '场景：%s\n%s%s' % (scene or '（未描述）', ('\n' + '\n'.join(char_info)) if char_info else '', ('\n动作/光线细节：' + detail) if detail else '')) try: raw = _llm_chat(providers, '你是资深动漫分镜画师。', user, max_tokens=500) raw = (raw or '').strip().strip('"').strip("'") if raw and len(raw) > 8: visual = raw except Exception: visual = '' if not visual: # LLM 不可用兜底：中文场景 + 英文画风/质量词 visual = scene or ('角色「' + (chars[0].get('name', '') if chars else '') + '」' if chars else '') # 系统统一追加：画风包 + 质量词 + 构图（保证画风稳定） parts = [visual] parts.append('style: ' + style_en) parts.append(quality_en) parts.append('vertical 9:16 composition, no text, no watermark') return ', '.join([x for x in parts if x and x.strip()]) def storyboard_episode(ep, ch_title, idx, assets, providers, target_shots=4): """把一集剧情用 LLM 拆成 3~5 个分镜镜头，返回 [{scene,prompt,lines,move,chars}]""" chars = assets.get('characters', []) char_names = [c.get('name', '') for c in chars if c.get('name')] system = '你是漫剧资深分镜师。把给定的小说/剧本片段拆成一组连贯的分镜镜头。' user = ('章节：%s，第%d集\n剧情片段：\n%s\n\n' '出场角色池：%s\n\n' '请只输出 JSON：{"shots":[{"scene":"画面描述（视觉化，含人物动作/景别/光线/情绪，可直接用于AI生图）",' '"prompt":"画面补充（构图/氛围/细节，可为空字符串）",' '"lines":"本镜头台词或旁白，格式：旁白：xxx 或 角色名：xxx；无台词写：无",' '"move":"推镜/拉镜/摇镜/静止",' '"chars":["出场角色名"]}]}\n' '要求：拆成 3~5 个镜头；镜头顺序与剧情连贯；chars 只从角色池里选；' 'lines 口语化带情绪；move 四选一；只输出 JSON。' % (ch_title, idx, ep[:6000], ', '.join(char_names) or '（无）')) raw = _llm_chat(providers, system, user, max_tokens=3500) d = _extract_json(raw) or {} shots = [] for sh in (d.get('shots') or []): if not isinstance(sh, dict): continue scene = str(sh.get('scene', '')).strip() if not scene: continue shots.append({ 'scene': scene, 'prompt': str(sh.get('prompt', '')).strip(), 'lines': str(sh.get('lines', '')).strip(), 'move': str(sh.get('move', '')).strip() or '静止', 'chars': [str(x) for x in (sh.get('chars') or []) if str(x).strip()], }) return shots def storyboard_all(chapters, episodes_map, assets, providers=None): """给所有集生成分镜，返回 {(章序号, 集序号): [shots]}""" if providers is None: providers = _load_llm_config() board = {} if not providers: return board for i, ch in enumerate(chapters): for j in range(len(episodes_map.get(i, []))): ep = episodes_map[i][j] try: board[(i, j)] = storyboard_episode(ep, ch['title'], j + 1, assets, providers) except Exception: board[(i, j)] = [] return board def extract_assets_heuristic(chapters): """规则兜底：对话归属+词频提人物；地点后缀提场景；量词提道具""" full = '\n'.join(c['content'] for c in chapters) chars = {} for m in re.finditer(r'([\u4e00-\u9fa5]{1,4})(?:说|道|问|喊|答|心想|低语|叹|喝|惊|怒道|笑道|沉声道|淡淡道)', full): n = m.group(1) if len(n) >= 2 and n not in ('他说', '她说', '众人', '大家', '其中'): chars[n] = chars.get(n, 0) + 1 names = [n for n, c in chars.items() if c >= 3][:12] scenes = [] for m in re.finditer(r'(?:来到|走进|回到|离开|站在|位于|在|赶往|逃到)([^，。！？；]{0,10}?)(?:的)?(?:山|城|府|殿|宫|院|镇|村|洞|林|谷|塔|楼|阁|河|湖|门|庙|厅|房|屋|堡|营|寨|堂|场|地)', full): s = m.group(1).strip() if s and s not in scenes: scenes.append(s) props = [] for m in re.finditer(r'(?:一把|一柄|一枚|一颗|一包|一件|一瓶|一卷|一块|一顶|一身|一根|一册|一页|一个|一只|一支|一张|一条|一盏|一份)([^，。！？\s]{1,6})', full): p = m.group(1).strip() if p and p not in props: props.append(p) return { 'title': os.path.basename(sys.argv[1]).rsplit('.', 1)[0] if len(sys.argv) > 1 else '未命名剧本', 'characters': [{'name': n, 'role': '', 'description': ''} for n in names], 'scenes': scenes[:15], 'props': props[:10], } # ---------------- 4) 生成画布 JSON ---------------- def build_canvas_json(title, chapters, episodes_map, assets, board=None): board = board or {} cards = [] uid = [1000] def cid(): uid[0] += 1 return 'c' + str(uid[0]) def note(x, y, text): return {'id': cid(), 'type': 'note', 'x': x, 'y': y, 'status': 0, 'data': {'text': text}} def scene(x, y, ep, shot): return {'id': cid(), 'type': 'scene', 'x': x, 'y': y, 'status': 0, 'data': {'ep': ep, 'scene': shot.get('scene', ''), 'charIds': [name_id[n] for n in shot.get('chars', []) if n in name_id], 'prompt': shot.get('prompt', ''), 'lines': shot.get('lines', ''), 'move': shot.get('move', '静止')}} def char(x, y, c): return {'id': cid(), 'type': 'char', 'x': x, 'y': y, 'status': 0, 'data': {'name': c.get('name', '角色'), 'appearance': c.get('description', ''), 'outfit': c.get('outfit', ''), 'feature': c.get('feature', ''), 'refImg': '', 'refNote': '由剧本导入提取，可再补一张参考图更稳'}} def tool(x, y, t, b): return {'id': cid(), 'type': 'tool', 'x': x, 'y': y, 'status': 0, 'data': {'title': t, 'body': b}} total_eps = sum(len(v) for v in episodes_map.values()) total_shots = 0 for ci in range(len(chapters)): for ej in range(len(episodes_map.get(ci, []))): total_shots += max(len(board.get((ci, ej), [])), 1) cards.append(note(40, 40, '📖 剧本导入总览\n\n《%s》\n章节：%d 章\n集数：%d 集\n分镜镜头：约 %d 个\n人物：%d 位\n场景：%d 个\n道具：%d 件\n\n' '■ 左侧：角色资产库（服装/特征已自动提取）\n■ 中间：每章一列，每集一排分镜镜头卡（已填画面/台词/运镜）\n■ 右上：场景/道具清单\n' '■ 直接点镜头卡「🎨 生图」出图，复制「台词」去配音' % (title, len(chapters), total_eps, total_shots, len(assets.get('characters', [])), len(assets.get('scenes', [])), len(assets.get('props', []))))) for i, c in enumerate(assets.get('characters', [])): cards.append(char(40, 320 + i * 360, c)) # 角色名 -> id 映射 name_id = {} for c in cards: if c['type'] == 'char': name_id[c['data']['name']] = c['id'] # 章节 + 分镜 x = 560 for ci, ch in enumerate(chapters): eps = episodes_map.get(ci, []) cards.append(note(x, 40, '📚 %s\n\n共拆为 %d 集：\n%s' % ( ch['title'], len(eps), '\n'.join('· 第%d集（%d个镜头）' % (j + 1, len(board.get((ci, j), [])) or 1) for j in range(len(eps)))))) for j in range(len(eps)): shots = board.get((ci, j)) or [{'scene': eps[j].strip()[:200], 'prompt': '', 'lines': '旁白：' + eps[j].strip()[:80], 'move': '静止', 'chars': []}] for k, shot in enumerate(shots): ep_title = '%s · 第%d集 · 镜头%d' % (ch['title'], j + 1, k + 1) cards.append(scene(x + 300 + j * 360, 40 + k * 360, ep_title, shot)) x += 300 + max(len(eps), 1) * 360 + 60 body = '【场景】\n' + '\n'.join('· ' + s for s in assets.get('scenes', [])) + '\n\n【道具】\n' + '\n'.join('· ' + p for p in assets.get('props', [])) cards.append(tool(x, 40, '🧰 场景 / 道具资产', body)) return {'cards': cards, 'view': {'tx': 0, 'ty': 0, 'z': 1}} # ---------------- 主入口 ---------------- def run_import(text, target_chars=1500, use_llm=True): chapters = split_chapters(text) episodes_map = {} for i, ch in enumerate(chapters): episodes_map[i] = split_episodes(ch['content'], target_chars) assets = None if use_llm: providers = _load_llm_config() if providers: try: assets = extract_assets_llm(chapters, providers) except Exception: assets = None if assets is None: assets = extract_assets_heuristic(chapters) board = {} providers = _load_llm_config() if use_llm and providers: board = storyboard_all(chapters, episodes_map, assets, providers) canvas = build_canvas_json(assets.get('title', '未命名剧本'), chapters, episodes_map, assets, board) shots_summary = {} for i, ch in enumerate(chapters): for j in range(len(episodes_map.get(i, []))): shots_summary['%s·第%d集' % (ch['title'], j + 1)] = len(board.get((i, j), [])) return { 'ok': True, 'title': assets.get('title', '未命名剧本'), 'chapters': [{'title': c['title'], 'episodes': len(episodes_map[i])} for i, c in enumerate(chapters)], 'assets': assets, 'shots': shots_summary, 'cards': canvas['cards'], 'view': canvas['view'], } def main(): args = sys.argv[1:] if not args: print(__doc__) sys.exit(1) path = args[0] out = 'aicomic-import.json' target = 1500 use_llm = True if '-o' in args: out = args[args.index('-o') + 1] if '--target' in args: target = int(args[args.index('--target') + 1]) if '--no-llm' in args: use_llm = False if not os.path.exists(path): print('文件不存在: %s' % path) sys.exit(1) text = open(path, encoding='utf-8').read() res = run_import(text, target_chars=target, use_llm=use_llm) with open(out, 'w', encoding='utf-8') as f: json.dump({'cards': res['cards'], 'view': res['view']}, f, ensure_ascii=False, indent=2) print('✅ 导入完成') print(' 标题:', res['title']) print(' 章节:', len(res['chapters']), '章 / 集数:', sum(c['episodes'] for c in res['chapters'])) print(' 人物:', len(res['assets'].get('characters', [])), '位 ->', ', '.join(c['name'] for c in res['assets'].get('characters', []))[:80]) print(' 场景:', len(res['assets'].get('scenes', [])), '个 / 道具:', len(res['assets'].get('props', []))) print(' 输出:', os.path.abspath(out)) print(' 提示: 打开画布 → 点「⬆ 导入」选这个 json') if __name__ == '__main__': main()