Seedance 短视频工程化：从单次生成到可复用视频流水线

import json
from openai import OpenAI

client = OpenAI(api_key="sk-xxx", base_url="https://api.deepseek.com/v1")

def build_storyboard(script: str) -> list[dict]:
    """把剧本拆成分镜，每个镜头只描述静态画面"""
    prompt = f"""把以下剧本拆成 3-5 个分镜。
每个分镜输出：{{"shot_id":int,"scene":str,"characters":list,"camera":str}}
不要描述动作过程，只描述这一帧的静态画面。
剧本：{script}
输出 JSON：{{"shots":[...]}}"""
    resp = client.chat.completions.create(
        model="deepseek-chat",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.3,
        response_format={"type": "json_object"})
    return json.loads(resp.choices[0].message.content)["shots"]

def validate_shots(shots: list[dict]) -> list[dict]:
    """检查分镜：人物是否一致、场景是否连贯"""
    issues = []
    chars_seen = set()
    for s in shots:
        for c in s["characters"]:
            if c in chars_seen:
                continue
            chars_seen.add(c)
        if not s.get("camera"):
            issues.append(f"镜头 {s['shot_id']} 缺少运镜描述")
    return issues

def generate_first_frame(shot: dict, style_ref: str) -> str:
    """调图像生成模型出首帧，返回图片URL"""
    prompt = f"""{style_ref}
场景：{shot['scene']}
人物：{', '.join(shot['characters'])}
镜头：{shot['camera']}
静态画面，不要动作，不要文字。"""
    resp = client.images.generate(
        model="seedream",
        prompt=prompt,
        size="1024x1024",
        n=1)
    return resp.data[0].url

def check_consistency(frames: list[dict]) -> dict:
    """用多模态模型对比多张首帧，检查人物一致性"""
    prompt = """对比以下多张图片中的人物。
判断：同一个人物在不同图片中的长相、服装、发型是否一致？
输出 JSON：{"consistent":bool,"issues":[str]}"""
    resp = client.chat.completions.create(
        model="deepseek-vl",
        messages=[{"role": "user", "content": [
            {"type": "text", "text": prompt},
            *[{"type": "image_url",
               "image_url": {"url": f["url"]}} for f in frames]
        ]}],
        response_format={"type": "json_object"})
    return json.loads(resp.choices[0].message.content)

def generate_clip(frame_url: str, motion: str,
                  max_retries: int = 2) -> dict:
    """用首帧生成视频片段，带重试上限"""
    for attempt in range(max_retries + 1):
        resp = client.video.generate(
            model="seedance",
            image_url=frame_url,
            prompt=motion,
            duration=5,
            resolution="720p")
        clip = resp.data[0]
        if clip.get("status") == "success":
            return {"url": clip["url"], "attempts": attempt + 1}
    return {"error": "生成失败", "attempts": max_retries + 1}