本地 LLM 实现男频 / 女频短剧剧本生成与结构化导出｜对接 AI 漫剧流水线工程实践

本地Qwen大模型 + 男女频编剧Skill模板 → 原始短剧剧本markdown输出
→ Python脚本：剧本清洗、格式校验、角色抽取
→ 转换为标准化分镜JSON → ComfyUI/Wan2.2批量渲染漫剧视频
→ 台词提取送入本地TTS → FFmpeg音视频混流、字幕压制输出成片

# 核心本地大模型推理库，用于加载GGUF量化模型
pip install llama‑cpp‑python pydantic re json -i https://pypi.tuna.tsinghua.edu.cn/simple

你是专业女频竖屏短剧编剧，严格遵守短剧创作规则：
1.每一集开篇3‑5秒必须抛出情绪钩子：背叛、误会、羞辱、情感危机；
2.单集60‑90秒，对话优先，减少大段心理描写，把心理活动转为动作神态；
3.单集中段设置小反转，结尾必须留下悬念钩子，引导看下一集；
4.输出剧本格式：【集名】、场景、人物、动作（括号）、台词；
5.输出不要多余解释，只输出剧本；
6.输出要适合9:16竖屏画面生成，动作描写具备画面可渲染性。

你是专业男频竖屏短剧编剧，严格遵守短剧创作规则：
1.开篇立刻制造绝境、被欺辱、身份被蔑视的钩子；
2.单集60‑90秒，爽点密集，短句对话，侧重对峙、气场、冲突；
3.中段安排小打脸反转，结尾留下身份悬念或者危机钩子；
4.输出剧本格式：【集名】、场景、人物、动作（括号）、台词；
5.禁止大段内心独白，全部转化为可视的动作行为；
6.输出适配9:16竖屏AI视频生成，所有描述必须可以转为画面。

from pydantic import BaseModel, Field
from typing import List, Optional
class SingleShot(BaseModel):
    """单场镜头结构，对接漫剧渲染"""
    scene_title: str = Field(description="场景名称，内/外，日/夜，地点")
    character_list: List[str] = Field(description="本场出场角色")
    action_desc: str = Field(description="动作神态描述，用于生成画面提示词")
    dialogue: str = Field(description="角色对话，用于TTS配音")
    hook_tag: Optional[str] = Field(description="钩子标记：开篇钩子 / 中段反转 / 结尾悬念")
class DramaEpisode(BaseModel):
    """单集短剧"""
    episode_no: int
    episode_title: str
    genre_tag: str = Field(description="男频 / 女频")
    shot_list: List[SingleShot]
class FullDramaScript(BaseModel):
    series_name: str
    character_setting: dict
    episodes: List[DramaEpisode]

from llama_cpp import Llama
import os
class LocalDramaLLM:
    def __init__(self, model_path: str, skill_system_prompt_path: str):
        """
        :param model_path: Qwen2.5‑7B‑Instruct‑Q4_K_M.gguf模型路径
        :param skill_system_prompt_path: skill_male.txt / skill_female.txt
        """
        self.llm = Llama(
            model_path=model_path,
            n_ctx=8192,
            n_gpu_layers=-1, # 全部层卸载到GPU，8G显卡Q4模型可用
            temperature=0.72,
            top_p=0.8,
            max_tokens=1800
        )
        with open(skill_system_prompt_path, "r", encoding="utf‑8") as f:
            self.system_prompt = f.read()
    def generate_episode_script(self, user_query: str) -> str:
        """输入创作要求，返回原始剧本文本"""
        messages = [
            {"role":"system","content":self.system_prompt},
            {"role":"user","content": user_query}
        ]
        output = self.llm.create_chat_completion(messages=messages)
        resp_text = output["choices"][0]["message"]["content"]
        return resp_text
if __name__ == "__main__":
    # =========切换男女频只需要修改skill路径=========
    skill_file = "./skill_female.txt"
    llm_gen = LocalDramaLLM(
        model_path="./models/qwen2.5‑7b‑instruct‑q4_k_m.gguf",
        skill_system_prompt_path=skill_file
    )
    # 用户输入创作需求
    req = "写女频重生短剧第1集，女主重生回到订婚宴被羞辱的时刻"
    raw_script = llm_gen.generate_episode_script(req)
    with open("./raw_drama.txt","w",encoding="utf‑8") as fw:
        fw.write(raw_script)
    print("原始剧本生成完成，保存 raw_drama.txt")
    print(raw_script)

python local_llm_drama_gen.py

import re
import json
from drama_schema import FullDramaScript,DramaEpisode,SingleShot
class DramaTextParser:
    def __init__(self, genre_tag="女频"):
        self.genre_tag = genre_tag
        # 匹配角色台词 【角色名】(动作)台词
        self.pat_dialogue = re.compile(r"【(.*?)】\s*（(.*?)）\s*(.*?)(?=【|$)", re.S)
        self.pat_scene_head = re.compile(r"第(\d+)场\s*(.*?)，(内|外)，(日|夜)")
    def parse_raw_text(self, raw_text:str, ep_no:int, ep_title:str):
        shot_result = []
        scene_blocks = re.split(r"第\d+场", raw_text)
        for block in scene_blocks:
            if len(block.strip()) <10:
                continue
            scene_match = self.pat_scene_head.search(block)
            scene_name = scene_match.group(0) if scene_match else "通用场景"
            dialog_items = self.pat_dialogue.findall(block)
            char_list = []
            action_all = []
            dialogue_all = []
            for char_name,action,words in dialog_items:
                char_list.append(char_name.strip())
                action_all.append(action.strip())
                dialogue_all.append(f"{char_name.strip()}：{words.strip()}")
            shot_item = SingleShot(
                scene_title=scene_name,
                character_list=list(set(char_list)),
                action_desc="；".join(action_all),
                dialogue="；".join(dialogue_all),
                hook_tag="auto"
            )
            shot_result.append(shot_item.model_dump())
        ep = DramaEpisode(
            episode_no=ep_no,
            episode_title=ep_title,
            genre_tag=self.genre_tag,
            shot_list=shot_result
        )
        return ep.model_dump()
if __name__ == "__main__":
    with open("./raw_drama.txt","r",encoding="utf‑8") as f:
        raw = f.read()
    parser = DramaTextParser(genre_tag="女频")
    ep_data = parser.parse_raw_text(raw,ep_no=1,ep_title="订婚宴的重生")
    out_json = {"episodes":[ep_data]}
    with open("./drama_struct.json","w",encoding="utf‑8") as fw:
        json.dump(out_json,fw,ensure_ascii=False,indent=2)
    print("结构化剧本输出 drama_struct.json")

python parse_drama_script.py

import json
def convert_drama_to_scene(drama_json_path, out_scene_json, style_tag="女频漫剧"):
    with open(drama_json_path,"r",encoding="utf‑8") as f:
        drama_data = json.load(f)
    scene_output = []
    for ep in drama_data["episodes"]:
        for idx,shot in enumerate(ep["shot_list"]):
            action_text = shot["action_desc"]
            scene_env = shot["scene_title"]
            pos_prompt = f"9:16竖屏，{style_tag}，{scene_env}，{action_text}，人物面部稳定，画风统一，镜头稳定，柔和光影"
            neg_prompt = "五官崩坏，肢体畸形，闪烁，画风突变，水印，文字，畸形手指"
            scene_output.append({
                "scene_id":f"ep{ep['episode_no']}_shot{idx+1}",
                "scene_description":f"{scene_env} {action_text}",
                "dialogue": shot["dialogue"],
                "positive_prompt":pos_prompt,
                "negative_prompt":neg_prompt,
                "frame_count":24
            })
    with open(out_scene_json,"w",encoding="utf‑8") as fw:
        json.dump(scene_output,fw,ensure_ascii=False,indent=2)
    print(f"漫剧分镜输出 {out_scene_json}，可送入批量渲染流水线")
if __name__ == "__main__":
    convert_drama_to_scene("./drama_struct.json","./scene_list.json",style_tag="女频漫剧")

python drama_to_scenelist.py

import json
def export_tts_text(scene_json, out_txt):
    with open(scene_json,"r",encoding="utf‑8") as f:
        data = json.load(f)
    lines = []
    for shot in data:
        dlg = shot.get("dialogue","")
        if dlg:
            lines.append(dlg)
    with open(out_txt,"w",encoding="utf‑8") as fw:
        fw.write("\n".join(lines))
    print(f"TTS台词文件输出 {out_txt}")
if __name__ == "__main__":
    export_tts_text("./scene_list.json","./tts_input.txt")

ffmpeg -f concat -safe 0 -i filelist.txt -c copy ep1_raw.mp4

ffmpeg -i ep1_raw.mp4 -vf "subtitles=subtitle.srt:force_style='FontName=SimHei,Fontsize=26,PrimaryColour=&H00ffffff&,Outline=2'" -c:a copy ep1_final.mp4

ffmpeg -i ep1_final.mp4 -c:v libx264 -preset medium -crf 23 -vf scale=408:720 -c:a aac -b:a 128k ep1_export.mp4