男女频短剧编剧 Skill 落地实践｜剧本自动流转到 AI 漫剧制作全链路工程方案

【103】编剧Skill模板 + 本地LLM推理 → 标准化短剧剧本
→ 剧本校验&清洗脚本 → 角色、场景、动作、台词字段抽取
→ 分镜自动生成模块 → 生成ComfyUI可读取的批量渲染JSON
→ IPAdapter角色特征绑定 → 画面一致性控制（AI漫剧渲染）
→ 台词分发本地TTS生成配音音频
→ FFmpeg音视频混流、字幕压制、成片导出

# 基础推理、数据解析依赖
pip install llama-cpp-python pydantic regex -i https://pypi.tuna.tsinghua.edu.cn/simple
# 图像处理、IP特征缓存辅助库
pip install pillow numpy

from llama_cpp import Llama
import os
class SkillDramaGenerator:
    def __init__(self, gguf_model_path: str):
        self.llm = Llama(
            model_path=gguf_model_path,
            n_ctx=8192,
            n_gpu_layers=-1,
            temperature=0.7,
            top_p=0.82,
            max_tokens=2000
        )
    def load_skill_prompt(self, skill_file_path: str):
        with open(skill_file_path, "r", encoding="utf-8") as f:
            return f.read()
    def generate_script(self, skill_prompt: str, user_requirement: str):
        msg_list = [
            {"role": "system", "content": skill_prompt},
            {"role": "user", "content": user_requirement}
        ]
        result = self.llm.create_chat_completion(messages=msg_list)
        script_content = result["choices"][0]["message"]["content"]
        return script_content
if __name__ == "__main__":
    generator = SkillDramaGenerator("./models/qwen2.5-7b-instruct-q4_k_m.gguf")
    # 切换：skill_male.txt 男频 / skill_female.txt 女频
    skill_prompt = generator.load_skill_prompt("./skill_female.txt")
    req = "写女频重生短剧第2集，女主当众拆穿继妹的伪装，结尾留下被男主误会的悬念"
    script = generator.generate_script(skill_prompt, req)
    with open("./drama_raw.txt", "w", encoding="utf-8") as f:
        f.write(script)
    print("剧本生成完成，准备进入AI漫剧制作分镜解析环节")
    print(script)

python skill_drama_entry.py

import re
import json
class DramaStoryboardConvert:
    def __init__(self):
        # 角色+动作+台词正则
        self.re_line = re.compile(r"【(.+?)】\s*（(.*?)）\s*(.+?)(?=【|$)", re.S)
        self.re_scene = re.compile(r"第(\d+)场\s*(.*?)，(内|外)，(日|夜)")
    def clean_text(self, raw_text: str):
        # 清洗markdown标记、空行
        clean = re.sub(r"#|```", "", raw_text)
        clean = re.sub(r"\n+", "\n", clean)
        return clean.strip()
    def parse_to_storyboard(self, raw_text: str, ep_name: str, genre: str):
        clean_txt = self.clean_text(raw_text)
        blocks = re.split(r"第\d+场", clean_txt)
        storyboard_list = []
        shot_index = 1
        for block in blocks:
            block = block.strip()
            if len(block) < 8:
                continue
            scene_match = self.re_scene.search(block)
            scene_info = scene_match.group(0) if scene_match else "通用场景"
            matches = self.re_line.findall(block)
            action_merge = []
            dialogue_merge = []
            char_set = set()
            for char_name, action, words in matches:
                char_set.add(char_name.strip())
                action_merge.append(action.strip())
                dialogue_merge.append(f"{char_name.strip()}：{words.strip()}")
            action_all = "，".join(action_merge)
            dialogue_all = "，".join(dialogue_merge)
            # 组装漫剧渲染提示词
            pos_prompt = f"9:16竖屏，短剧漫剧，{scene_info}，{action_all}，人物五官稳定，画风统一，光影柔和，镜头平稳"
            neg_prompt = "画面闪烁，人物变形，多手指，水印，文字，画风跳变，畸形肢体"
            storyboard_list.append({
                "shot_id": f"{ep_name}_shot{shot_index}",
                "scene": scene_info,
                "characters": list(char_set),
                "action": action_all,
                "dialogue": dialogue_all,
                "positive": pos_prompt,
                "negative": neg_prompt
            })
            shot_index += 1
        return storyboard_list
if __name__ == "__main__":
    with open("./drama_raw.txt", "r", encoding="utf-8") as f:
        raw = f.read()
    converter = DramaStoryboardConvert()
    res = converter.parse_to_storyboard(raw, ep_name="ep02", genre="女频")
    with open("./comfy_storyboard.json", "w", encoding="utf-8") as fw:
        json.dump(res, fw, ensure_ascii=False, indent=2)
    print("分镜JSON已生成，可直接导入AI漫剧批量渲染工作流")

python script_to_comfy_storyboard.py

import requests
import json
import time
COMFY_URL = "http://127.0.0.1:8188"
def queue_prompt(workflow_json):
    payload = {"prompt": workflow_json}
    resp = requests.post(f"{COMFY_URL}/prompt", json=payload)
    return resp.json()
def batch_render(storyboard_path, workflow_template_path):
    with open(storyboard_path, "r", encoding="utf-8") as f:
        shots = json.load(f)
    with open(workflow_template_path, "r", encoding="utf-8") as f:
        base_workflow = json.load(f)
    for shot in shots:
        # 替换工作流内提示词节点
        base_workflow["6"]["inputs"]["text"] = shot["positive"]
        base_workflow["7"]["inputs"]["text"] = shot["negative"]
        print(f"提交任务：{shot['shot_id']}")
        queue_prompt(base_workflow)
        time.sleep(3)
if __name__ == "__main__":
    batch_render("./comfy_storyboard.json", "./workflow_template.json")

import json
with open("./comfy_storyboard.json","r",encoding="utf-8") as f:
    data = json.load(f)
lines = []
for item in data:
    lines.append(item["dialogue"])
with open("./tts_input.txt","w",encoding="utf-8") as f:
    f.write("\n".join(lines))

# 拼接所有分镜视频
ffmpeg -f concat -safe 0 -i filelist.txt -c copy ep2_raw.mp4
# 嵌入字幕
ffmpeg -i ep2_raw.mp4 -vf "subtitles=sub.srt:force_style='FontName=SimHei,Fontsize=28,Outline=2'" -c:a copy ep2_sub.mp4
# 统一输出竖屏短剧标准参数
ffmpeg -i ep2_sub.mp4 -c:v libx264 -preset medium -crf 24 -vf scale=408:720 ep2_final.mp4