8G 显存 Windows 本地部署 AI 漫剧全链路方案：GGUF 模型调度、剧本分镜 Agent 自动化、竖屏成片合成实践

import torch
import psutil
import os
def hardware_check():
    info = {}
    # CPU内存检测
    mem = psutil.virtual_memory()
    info["total_ram_gb"] = round(mem.total / 1024**3, 2)
    info["available_ram_gb"] = round(mem.available / 1024**3, 2)
    # GPU显存检测
    if torch.cuda.is_available():
        gpu_count = torch.cuda.device_count()
        gpu_list = []
        for i in range(gpu_count):
            free, total = torch.cuda.mem_get_info(i)
            gpu_info = {
                "gpu_id": i,
                "name": torch.cuda.get_device_name(i),
                "total_vram_gb": round(total / 1024**3, 2),
                "free_vram_gb": round(free / 1024**3, 2)
            }
            gpu_list.append(gpu_info)
        info["gpu"] = gpu_list
    else:
        info["gpu"] = None
    # GGUF模型推荐策略
    suggest = ""
    if info["gpu"]:
        vram = info["gpu"][0]["total_vram_gb"]
        if vram <= 8:
            suggest = "推荐GGUF Q4_K_M量化模型，开启显存分层调度，权重溢出置换到内存；不建议Q6以上量化"
        elif 8 < vram <=12:
            suggest = "推荐GGUF Q5_K_M，可同时加载LLM+绘图模型"
        else:
            suggest = "可使用Q6_K/Q8_0，支持更大上下文批量分镜处理"
    else:
        suggest = "无CUDA显卡，仅CPU推理GGUF，速度很慢，不适合AI漫剧批量生产"
    info["model_suggest"] = suggest
    return info
if __name__ == "__main__":
    res = hardware_check()
    print("===== 硬件检测报告 =====")
    for k,v in res.items():
        print(f"{k}: {v}")

from llama_cpp import Llama
import json
# 本地部署GGUF模型路径，自行替换为HuggingFace下载后的模型文件
llm = Llama(
    model_path="./models/llama3-8b.Q4_K_M.gguf",
    n_ctx=4096,
    n_gpu_layers=20, # 8G卡设置20层 offload到GPU，剩余在内存
    verbose=False
)
prompt_template = """
你是AI漫剧分镜Agent，任务：把下面的短剧剧本拆成适合9:16竖屏AI绘图的分镜。
输出严格JSON数组，不要额外解释。
每条分镜字段：
shot_id:镜头编号
scene_desc:场景描述
pos_prompt:绘图正向提示词，适合漫剧，写实，竖屏短视频
neg_prompt:反向提示词，畸形、扭曲、多手指、模糊、低画质
duration:镜头时长（秒）
ref_img_path:角色参考图固定路径./ref/char01.png
剧本内容：
{script_text}
"""
def parse_script_to_shot(script_path:str):
    with open(script_path,"r",encoding="utf-8") as f:
        script_text = f.read()
    prompt = prompt_template.format(script_text=script_text)
    output = llm(
        prompt=prompt,
        max_tokens=2048,
        temperature=0.6,
        stop=["```"]
    )
    raw = output["choices"][0]["text"]
    shots = json.loads(raw)
    # 保存分镜清单
    with open("./output/shot_list.json","w",encoding="utf-8") as fw:
        json.dump(shots,fw,ensure_ascii=False,indent=2)
    print(f"生成{len(shots)}条分镜，已保存 shot_list.json")
    return shots
if __name__ == "__main__":
    parse_script_to_shot("./script/drama.txt")

ffmpeg -i shot_01.mp4 -vf "subtitles=shot_01.srt:force_style='FontName=Microsoft YaHei,FontSize=30,PrimaryColour=&HFFFFFF,OutlineColour=&H40000000'" -c:a copy shot_01_sub.mp4

file 'shot_01_sub.mp4'
file 'shot_02_sub.mp4'
file 'shot_03_sub.mp4'

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

ffmpeg -i raw.mp4 -vf "scale=720:1280,fps=24" -b:v 1500k -c:a aac final_916.mp4