AI 视频制作全流程：从脚本到成片

SCRIPT_GENERATION_PROMPT = """
你是一位专业编剧。请基于以下要求生成一个视频脚本：

主题：{topic}
时长：{duration}秒
风格：{style}  # 纪录片/剧情/广告/MV
目标受众：{audience}
情绪基调：{mood}  # 振奋/温暖/悬疑/幽默

输出格式：
1. 一句话概要（Logline）：25字以内
2. 故事大纲：100字以内
3. 详细分幕脚本：
   - 每幕包含：场景描述、旁白/对白、视觉风格提示、预计时长
4. 视觉关键词列表：用于后续 AI 视频生成的风格参考

约束：
- 每个场景的视觉描述必须是具体的、可生成的画面
- 避免需要复杂人物交互的场景（当前 AI 视频的弱项）
- 优先考虑视觉表现力强的场景
"""

{
  "title": "城市晨曦",
  "logline": "一座城市从沉睡到苏醒的三分钟视觉诗篇",
  "total_duration": 180,
  "scenes": [
    {
      "id": 1,
      "duration": 15,
      "location": "城市天际线",
      "time_of_day": "黎明前",
      "shot_type": "广角远景",
      "visual_description": "深蓝色天空中星光渐隐，城市轮廓在薄雾中若隐若现，远处建筑灯光零星闪烁",
      "narration": "在第一缕光到来之前，城市还在梦中",
      "style_keywords": ["cinematic", "aerial", "blue hour", "misty", "wide angle"],
      "negative_prompts": ["people", "text", "watermark"]
    },
    {
      "id": 2,
      "duration": 12,
      "location": "街道",
      "time_of_day": "日出",
      "shot_type": "中景",
      "visual_description": "空旷的街道上，第一缕阳光穿过建筑间隙，形成金色光带，落叶微微飘动",
      "narration": "光，是城市最早的访客",
      "style_keywords": ["golden hour", "empty street", "light rays", "cinematic", "slow motion"],
      "negative_prompts": ["crowd", "vehicles", "text"]
    }
  ]
}

工作流：
1. 生成主角参考图 → 固定角色外观
2. 生成场景概念图 → 固定环境风格
3. 生成色彩脚本 → 固定调色方案
4. 将参考图作为视频生成的输入条件

class StyleConsistencyManager:
    def __init__(self):
        self.style_seeds = {}
        self.reference_images = {}
    
    def register_style(self, style_name, seed, reference_image):
        """注册一个风格及其参考图"""
        self.style_seeds[style_name] = seed
        self.reference_images[style_name] = reference_image
    
    def generate_with_consistency(self, scene, style_name):
        """使用注册的风格生成视频"""
        return video_generator.generate(
            prompt=scene.visual_description,
            reference_image=self.reference_images[style_name],
            seed=self.style_seeds[style_name],
            duration=scene.duration,
            aspect_ratio="16:9"
        )

class VideoGenerationConfig:
    """视频生成的核心参数配置"""
    
    def __init__(self):
        # 基本参数
        self.duration = 15          # 秒
        self.fps = 30
        self.aspect_ratio = "16:9"
        self.resolution = "1920x1080"
        
        # 视觉控制
        self.motion_strength = 0.7  # 运动幅度 0-1
        self.camera_movement = "slow_pan_left"  # 运镜方式
        self.style_intensity = 0.8  # 风格强度
        
        # 一致性控制
        self.reference_image = None  # 参考图
        self.consistency_strength = 0.85  # 一致性强度
        
        # 负面提示
        self.negative_prompt = "blurry, low quality, distorted, watermark, text, ugly"

async def generate_scene_variants(scene, n=4):
    """为一个场景生成多个候选版本"""
    tasks = []
    for i in range(n):
        config = VideoGenerationConfig()
        config.motion_strength = 0.5 + i * 0.1  # 变化运动幅度
        tasks.append(generate_single(scene, config))
    
    variants = await asyncio.gather(*tasks)
    
    # 使用视觉模型评分排序
    scored = []
    for variant in variants:
        score = await evaluate_video_quality(variant, scene)
        scored.append((variant, score))
    
    scored.sort(key=lambda x: x[1], reverse=True)
    return scored[0][0]  # 返回最佳版本

from moviepy.editor import VideoFileClip, concatenate_videoclips, AudioFileClip, CompositeVideoClip

def assemble_final_video(clips_data, narration_audio, bgm_audio):
    """
    组装最终视频
    clips_data: [{path, start, end, transition_in, transition_out}]
    """
    clips = []
    for cd in clips_data:
        clip = VideoFileClip(cd['path']).subclip(cd['start'], cd['end'])
        
        # 添加转场效果
        if cd.get('transition_in') == 'fade':
            clip = clip.fadein(0.5)
        if cd.get('transition_out') == 'fade':
            clip = clip.fadeout(0.5)
        
        clips.append(clip)
    
    # 拼接视频
    final_video = concatenate_videoclips(clips, method='compose')
    
    # 添加旁白
    narration = AudioFileClip(narration_audio)
    # 添加背景音乐（降低音量）
    bgm = AudioFileClip(bgm_audio).volumex(0.15)
    
    # 混音
    from moviepy.audio.AudioClip import CompositeAudioClip
    mixed_audio = CompositeAudioClip([narration, bgm])
    final_video = final_video.set_audio(mixed_audio)
    
    return final_video

def color_grade_unify(clips, target_palette):
    """
    将所有剪辑统一到目标色彩风格
    """
    graded_clips = []
    for clip in clips:
        # 计算 LUT (Look-Up Table)
        lut = generate_lut(clip, target_palette)
        # 应用 LUT
        graded = apply_lut(clip, lut, strength=0.7)
        graded_clips.append(graded)
    
    return graded_clips

class VideoQualityAssessor:
    """AI 视频质量自动评估系统"""
    
    def assess(self, video_path, script):
        report = {}
        
        # 1. 技术质量
        report['technical'] = {
            'resolution': check_resolution(video_path),
            'bitrate': check_bitrate(video_path),
            'fps_stability': check_fps(video_path),
            'artifacts': detect_artifacts(video_path)  # AI 生成伪影
        }
        
        # 2. 内容一致性
        report['consistency'] = {
            'style_consistency': check_style_consistency(video_path, script),
            'character_consistency': check_character_consistency(video_path, script),
            'color_consistency': check_color_consistency(video_path)
        }
        
        # 3. 脚本忠实度
        report['fidelity'] = {
            'scene_coverage': check_scene_coverage(video_path, script),
            'narration_match': check_narration_alignment(video_path, script),
            'duration_match': check_duration(video_path, script)
        }
        
        # 4. 观感质量
        report['aesthetic'] = {
            'composition_score': evaluate_composition(video_path),
            'motion_smoothness': evaluate_motion(video_path),
            'visual_appeal': evaluate_visual_appeal(video_path)
        }
        
        return report

async def fix_character_consistency(video_path, reference_face):
    """
    使用 face-swap 技术统一角色外观
    """
    frames = extract_frames(video_path)
    fixed_frames = []
    
    for frame in frames:
        if detect_face(frame):
            fixed = await face_swap(frame, reference_face, strength=0.6)
            fixed_frames.append(fixed)
        else:
            fixed_frames.append(frame)
    
    return frames_to_video(fixed_frames, fps=30)

OUTPUT_FORMATS = [
    {"name": "youtube", "resolution": "1920x1080", "fps": 30, "bitrate": "8M", "format": "mp4"},
    {"name": "douyin", "resolution": "1080x1920", "fps": 30, "bitrate": "6M", "format": "mp4"},
    {"name": "xiaohongshu", "resolution": "1080x1350", "fps": 30, "bitrate": "5M", "format": "mp4"},
    {"name": "web", "resolution": "1280x720", "fps": 24, "bitrate": "3M", "format": "webm"},
]

SEO_PROMPT = """
基于以下视频脚本，生成适合 {platform} 的发布元数据：

脚本摘要：{script_summary}
视频时长：{duration}
目标受众：{audience}

输出：
1. 标题（{platform}风格，吸引点击）
2. 描述（200字以内，包含关键信息）
3. 标签列表（10-15个相关标签）
4. 封面建议（描述最佳封面帧的画面）
"""