AI视频全流程生成深度实践：从剧本到成片的自动化流水线

用户创意 → 剧本扩写 → 角色/场景设定 → 分镜规划 → 
关键帧生成 → 视频分段生成 → 音频合成 → 后期拼接[reference:4]

from langgraph.graph import StateGraph
from langchain_openai import ChatOpenAI

class ScriptAgent:
    def __init__(self):
        self.llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.7)
    
    def generate_script(self, idea: str) -> list:
        prompt = f"""
        将以下创意扩写为5个分镜的短视频剧本。
        每个分镜包含：scene_id、description（画面描述）、dialogue（台词）、duration（秒）。
        创意：{idea}
        输出JSON数组。
        """
        response = self.llm.invoke(prompt)
        return json.loads(response.content)

# 基于ModelScope MCP的图像生成配置[reference:10]
{
    "service_type": "image_generation",
    "model": "damo/cv_diffusion_text2image",
    "params": {
        "prompt": "{scene_desc}",
        "negative_prompt": "文字,水印",
        "width": 1080,
        "height": 1920
    }
}

def enhance_prompt(scene_desc: str) -> str:
    base_prompt = "masterpiece, best quality, 8k"
    return f"{base_prompt}, {scene_desc} --lora_weights=animestyle_v2:0.7"

# 基于首帧生成视频
python scripts/generate_video.py \
    --prompt "Slow dolly forward, ambient evening sounds" \
    --image scene_01.png \
    --output clip_01.mp4

# 使用Seedance 2.0
python scripts/generate_video_seedance.py \
    --prompt "cinematic shot" \
    --image scene_01.png

from diffusers import AnimateDiffPipeline, MotionAdapter
from diffusers.utils import export_to_video

adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2")
pipe = AnimateDiffPipeline.from_pretrained(
    "SG161222/Realistic_Vision_V5.1_noVAE",
    motion_adapter=adapter,
    torch_dtype=torch.float16
).to("cuda")

output = pipe(
    prompt="a cute dog running in a park, cinematic, highly detailed",
    num_frames=24,
    guidance_scale=7.5
)
export_to_video(output.frames[0], "output.mp4", fps=8)

# 文生视频（FP8量化，降低显存）
python cli_demo_quantization.py \
    --prompt "A girl riding a bike." \
    --model_path THUDM/CogVideoX-2b \
    --quantization_scheme fp8

# 生成旁白
python scripts/generate_tts.py --text "The street remembers." --output vo.mp3

# 生成背景音乐
python scripts/generate_music.py --prompt "Cinematic ambient strings, slow"

import subprocess

def assemble_video(clips: list, audio: str, output: str):
    # 生成FFmpeg concat列表
    with open("concat_list.txt", "w") as f:
        for clip in clips:
            f.write(f"file '{clip}'\n")
    # 执行合成
    cmd = [
        "ffmpeg", "-f", "concat", "-safe", "0",
        "-i", "concat_list.txt",
        "-i", audio,
        "-c:v", "libx264", "-c:a", "aac",
        "-shortest", output
    ]
    subprocess.run(cmd, check=True)

from langgraph.graph import StateGraph, END

class VideoProductionGraph:
    def __init__(self):
        self.graph = StateGraph(dict)
        self.graph.add_node("script", self.generate_script)
        self.graph.add_node("storyboard", self.generate_storyboard)
        self.graph.add_node("video", self.generate_video)
        self.graph.add_node("audio", self.generate_audio)
        self.graph.add_node("assemble", self.assemble)
        # 定义流转
        self.graph.set_entry_point("script")
        self.graph.add_edge("script", "storyboard")
        self.graph.add_edge("storyboard", "video")
        self.graph.add_edge("video", "audio")
        self.graph.add_edge("audio", "assemble")
        self.graph.add_edge("assemble", END)
    
    def run(self, idea: str) -> str:
        return self.graph.compile().invoke({"idea": idea})