单 8G 显卡离线跑通 AI 漫剧长剧集，自动换场景不崩人设，全链路 Agent 无人值守渲染方案

@echo off
chcp 65001
set CUDA_VISIBLE_DEVICES=0
set PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:128
set COMFYUI_VRAM_MODE=lowvram
set COMFYUI_AUTO_UNLOAD=1
set COMFYUI_FP8=1
python main.py --enable-agent-api
pause

import json
import os
from ollama import chat
SCRIPT_PATH = "./script/episode01.txt"
SHOT_OUTPUT = "./shot_auto.json"
# 角色缓存映射表，角色名绑定特征文件
CHAR_MAP = {
    "女主": "./character_cache/girl.bin",
    "男主": "./character_cache/boy.bin",
    "配角": "./character_cache/support.bin"
}
prompt_template = """
你是短剧分镜工程师，输入剧本文本，输出标准JSON数组。
要求：
1. 一条对话/动作对应一条镜头；
2. 识别当前出场角色，填入角色名称；
3. 根据场景自动生成画面提示词，9:16竖屏漫剧，写实画风；
4. 输出时长（单镜头24帧）；
5. 只输出JSON，不要额外解释。
字段：shot_id,role_name,scene,prompt,duration
"""
def parse_script_to_shot(script_text):
    resp = chat(model="qwen3:8b", messages=[
        {"role":"system","content":prompt_template},
        {"role":"user","content":script_text}
    ])
    return resp["message"]["content"]
if __name__ == "__main__":
    with open(SCRIPT_PATH,"r",encoding="utf-8") as f:
        content = f.read()
    res = parse_script_to_shot(content)
    shot_data = json.loads(res)
    # 绑定角色特征路径
    for shot in shot_data:
        if shot["role_name"] in CHAR_MAP:
            shot["char_emb_path"] = CHAR_MAP[shot["role_name"]]
    with open(SHOT_OUTPUT,"w",encoding="utf-8") as f:
        json.dump(shot_data,f,ensure_ascii=False,indent=2)
    print(f"分镜自动生成完成，保存至 {SHOT_OUTPUT}")

import requests
import json
import os
import time
COMFY_API = "http://127.0.0.1:8188/prompt"
SHOT_FILE = "./shot_auto.json"
OUTPUT = "./output/episode01"
MAX_RETRY = 2
os.makedirs(OUTPUT,exist_ok=True)
def submit_prompt(workflow):
    try:
        r = requests.post(COMFY_API,json=workflow,timeout=300)
        return r.json()
    except Exception as e:
        print("任务提交异常：",e)
        return None
if __name__ == "__main__":
    with open(SHOT_FILE,"r",encoding="utf-8") as f:
        shot_list = json.load(f)
    for shot in shot_list:
        flag = os.path.join(OUTPUT,f"{shot['shot_id']}.ok")
        if os.path.exists(flag):
            print(f"镜头{shot['shot_id']}已完成，跳过")
            continue
        retry = 0
        success = False
        while retry <= MAX_RETRY and not success:
            print(f"开始渲染镜头{shot['shot_id']}，重试次数：{retry}")
            # 这里填入你预调好的Wan2.2工作流，替换prompt和角色embedding路径
            workflow_payload = {
                "prompt":{
                    "6":{"inputs":{"text":shot["prompt"]}},
                    "12":{"inputs":{"emb_path":shot["char_emb_path"]}}
                }
            }
            ret = submit_prompt(workflow_payload)
            if ret:
                open(flag,"w",encoding="utf-8").write("done")
                success = True
            else:
                retry +=1
                time.sleep(10)
        if not success:
            print(f"镜头{shot['shot_id']}多次渲染失败，标记待人工处理")
            open(os.path.join(OUTPUT,f"{shot['shot_id']}.fail"),"w")

from tts import local_tts
import srt
shot_json = "./shot_auto.json"
audio_dir = "./audio"
sub_list = []
os.makedirs(audio_dir,exist_ok=True)
with open(shot_json,"r",encoding="utf-8") as f:
    shots = json.load(f)
start = 0.0
for shot in shots:
    text = shot["prompt"]
    audio_path = os.path.join(audio_dir,f"{shot['shot_id']}.wav")
    local_tts.generate(text, audio_path)
    duration = shot["duration"] / 12
    sub = srt.Subtitle(index=shot["shot_id"],
                       start=srt.Timedelta(seconds=start),
                       end=srt.Timedelta(seconds=start + duration),
                       content=text)
    sub_list.append(sub)
    start += duration
srt_content = srt.compose(sub_list)
with open("./subtitle.srt","w",encoding="utf-8") as f:
    f.write(srt_content)

ffmpeg -y -f concat -safe 0 -i video_list.txt -filter_complex "amix=inputs=1[a];subtitles=subtitle.srt:force_style='Fontname=SimHei,Fontsize=26,PrimaryColour=&Hffffff&OutlineColour=&H000000&Outline=2'" -map 0:v -map "[a]" -c:v libx264 -preset medium -crf 24 -vf "scale=720:1280,fps=24" -c:a aac -b:a 128k episode_final.mp4

file './output/episode01/shot1.mp4'
file './output/episode01/shot2.mp4'
file './output/episode01/shot3.mp4'

import cv2
import glob
import os
def calc_blur(img):
    gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
    return cv2.Laplacian(gray, cv2.CV_64F).var()
if __name__ == "__main__":
    img_paths = glob.glob("./output/episode01/*.png")
    bad = []
    for p in img_paths:
        im = cv2.imread(p)
        score = calc_blur(im)
        if score < 100:
            bad.append(p)
    print("检测到画质异常镜头：")
    for b in bad:
        print(b)
        os.rename(b, b.replace(".png","_bad.png"))