Agent/knowledge_batch_20260409.json

{
  "upload_batch": "AI 图像/视频生成工作流研究",
  "upload_date": "2026-04-09",
  "source_type": "research",
  "knowledge_count": 6,
  "knowledge": [
    {
      "id": "knowledge-20260409-001",
      "task": "AI 角色一致性工作流：双参考图锁定 + 角色参考表生成",
      "content": "## 核心发现\n单一参考图会导致身份漂移，使用**双参考图锁定**是 100% 角色一致性的关键。\n\n## 完整工作流（Nano Banana Pro + Kling）\n\n### 步骤 1：生成角色参考表\n**输入**：双张参考图\n**工具**：Nano Banana Pro（Higgsfield AI）\n**输出**：角色参考表（Character Sheet）\n**提示词模板**：\n```\n[A side-by-side photograph combining a detailed portrait and a full body character design. On the left is a close-up portrait with noticeable details (sweat, expression). On the right is a full-body photograph facing forward. Background is neutral studio grey. Cinematic lighting.]\n```\n\n### 步骤 2：生成分镜接触表\n**输入**：角色参考表\n**工具**：Nano Banana Pro\n**输出**：9 格电影分镜\n**提示词模板**：\n```\nCreate a cinematic contact sheet grid containing a sequence of nine separate dynamic film stills.\nInclude a mix of shot types:\n- Wide angle landscape shots\n- Medium action shots\n- Extreme close-up shots focusing on face/hands/details\nAll images must have heavy film grain, dramatic low-key lighting.\n```\n\n### 步骤 3：升格至 4K\n**输入**：分镜帧\n**工具**：Nano Banana Pro / Topaz Gigapixel\n**输出**：4K 高分辨率图像\n\n### 步骤 4：视频生成\n**输入**：高分辨率帧\n**工具**：Kling 1.0（光照/天气变化）或 Kling 2.6（内置语音）\n**输出**：带运动/声音的视频\n\n## 替代方案：AI Video School 角色表提示词\n\n**有参考图**：\n```\nCreate a professional character reference sheet based strictly on the uploaded reference image.\nArrange into two horizontal rows:\n- Top row: four full-body views (front, left profile, right profile, back)\n- Bottom row: three close-up portraits (front, left profile, right profile)\nMaintain perfect identity consistency, A-pose, consistent scale, clean panel separation.\nLighting consistent across all panels.\n```\n\n**无参考图**：将上述提示词中\"uploaded reference image\"替换为具体角色描述\n\n## 关键参数\n- 布局：两行（上排全身 4 视角，下排特写 3 视角）或四列垂直布局\n- 姿势：A-pose（放松站姿）\n- 光照：所有面板一致的方向、强度、柔和度\n- 背景：干净中性纯色背景",
      "types": ["strategy"],
      "score": 5,
      "source_urls": ["YouTube: CMPhDepMQS4", "YouTube: f4HcdR3cd4M", "YouTube: 2psBexPkw3I"],
      "domain": "AI 图像生成",
      "tools": ["Nano Banana Pro", "Kling", "Higgsfield AI"],
      "intent": "角色一致性生成"
    },
    {
      "id": "knowledge-20260409-002",
      "task": "Stable Diffusion 局部细节放大：2 步放大技术（潜空间+SEEDVR）",
      "content": "## Z Image Turbo + 2 步放大工作流\n\n### 核心参数设置\n```\nKsampler Group 1 & 2: dpmpp_sde + beta\nUpscale method: Bislerp\nScale by (Upscale Latent by): 1.25\nDenoise (Group 2 Ksampler): 0.35\n```\n\n### 完整步骤\n\n| 步骤 | 输入 | 工具 | 输出 | 说明 |\n|------|------|------|------|------|\n| 1 | 原始图像 | Z Image Turbo | 照片级结果 | 快速生成基础图像 |\n| 2 | 潜空间 | Latent Upscale | 1.25×放大 | Bislerp 方法，不增加 VRAM |\n| 3 | 放大潜空间 | SEEDVR | 增强真实感 | 细节修复，保持 VRAM 效率 |\n\n### 模型下载\n- 主扩散模型：https://huggingface.co/Comfy-Org/z_im...\n- Text Encoder (Qwen 3 4B): https://huggingface.co/Comfy-Org/z_im...\n- VAE: https://huggingface.co/Comfy-Org/z_im...\n\n## 替代方案：SUPIR + SRPO 双阶段放大\n\n**第一阶段**：SUPIR 初始放大\n**第二阶段**：SRPO 皮肤纹理增强\n**适用场景**：去除 AI 塑料皮肤，恢复真实皮肤纹理\n\n## 视频放大方案：Wan 2.2 + Ultimate SD Upscaler\n- 自动分批处理长视频\n- 无缝拼接\n- 支持 2K/HD/4K 输出\n- 生成式放大（增加细节而非传统插值）\n\n## 避坑指南\n⚠️ 直接放大导致细节丢失 → 使用 ControlNet 保持结构\n⚠️ VRAM 不足 → 使用 SEEDVR 潜空间放大\n⚠️ 皮肤塑料感 → 使用 SUPIR+SRPO 双阶段修复",
      "types": ["strategy"],
      "score": 5,
      "source_urls": ["YouTube: rmp8z3svfco", "YouTube: Q-9Wbk_AX7c", "YouTube: NpNagmQI4yg"],
      "domain": "AI 图像放大",
      "tools": ["ComfyUI", "Z Image Turbo", "SEEDVR", "SUPIR", "SRPO"],
      "intent": "局部细节放大"
    },
    {
      "id": "knowledge-20260409-003",
      "task": "ComfyUI 面部一致性工作流：Qwen Edit 2511 + FaceDetailer",
      "content": "## 100% 面部一致性工作流\n\n### 工具链\n- ComfyUI\n- Qwen Edit 2511 模型\n- Face LoRA\n- SeedVR2 + SRPO（升格）\n\n### Face LoRA 下载\nhttps://modelscope.cn/models/DiffSynt...\n\n### 工作流 1：单头像到多肖像\n\n| 步骤 | 输入 | 工具 | 输出 | 参数 |\n|------|------|------|------|------|\n| 1 | 单张头像 | Qwen Edit 2511 | 多张肖像 | Lightning LoRA |\n| 2 | 参考图像×多张 | Qwen Edit | 面部一致性 | 多图像输入锁定身份 |\n| 3 | 生成图像 | Pose Control | 姿势控制 | |\n| 4 | 面部区域 | FaceDetailer | 面部细化 | FaceDetailer Pass |\n| 5 | 细化图像 | SeedVR2 | 升格输出 | |\n\n### 工作流 2：切换模型改变面部角度\n使用不同模型生成不同角度的面部，保持身份一致性\n\n## Flux Uncensored 全身/肖像工作流\n\n### LoRA 配置\n- LoRA 1: 皮肤增强\n- LoRA 2: 手部增强\n- 其他 LoRA: 细节增强\n\n### 工作流设置\n- Dual Clip Loader（新双剪辑加载器）\n- 工作流 1：i2i（图像到图像）设置\n- 工作流 2：t2i（文本到图像）直接生成\n\n## 关键发现\n- 多参考图输入是面部一致性的核心\n- FaceDetailer Pass 专门细化面部区域\n- Lightning LoRA 加速生成过程",
      "types": ["strategy"],
      "score": 5,
      "source_urls": ["YouTube: ydzMmHhm_9c", "YouTube: N7GbJ97vJow"],
      "domain": "AI 面部生成",
      "tools": ["ComfyUI", "Qwen Edit 2511", "FaceDetailer", "Flux"],
      "intent": "面部一致性生成"
    },
    {
      "id": "knowledge-20260409-004",
      "task": "视频换脸工作流：Nuke+WAN 2.1 与 LTX 2.3 对比",
      "content": "## 方案 1：Nuke + WAN 2.1 头部替换（专业级）\n\n### 完整步骤\n\n| 步骤 | 工具 | 操作 | 输出 |\n|------|------|------|------|\n| 1 | Nuke | 稳定化 + 绘制占位符 | 锁定镜头 |\n| 2 | ComfyUI + WAN 2.1 | 头部替换 | 合成面部序列 |\n| 3 | Nuke | Vitmat 遮罩 + 反向稳定化 | 最终合成 |\n\n### 核心技巧\n- **绘制占位符**是获得精确度的关键秘密\n- WAN 锁定到占位符，混合身份、表情和光照\n- 匹配运动模糊，保持微表情\n- 反向稳定化恢复原始运动\n\n### 项目文件\n- Benji AI Patreon: https://www.patreon.com/posts/comfyui...\n- CTRL_FX 项目文件：https://www.patreon.com/posts/nuke-wa...\n\n## 方案 2：LTX 2.3 视频换脸（简化版）\n\n### 工具\n- ComfyUI\n- LTX 2.3\n- BFS Node\n- Best Face Swap Video V3 LoRA\n\n### 模型下载\n- 模型：https://huggingface.co/Alissonerdx/BF...\n- BFSNodes: https://github.com/alisson-anjos/Comf...\n\n### 关键修复（必须执行）\n⚠️ **安装自定义调度器（9 步）**：否则面部会扭曲\n错误的采样器调度器会破坏生成\n\n### 测试结果\n- 面部相似度评分：70/100\n- ⚠️ **禁止跨种族换脸**（效果极差）\n- 极端测试：水下场景、突然场景切换、特写镜头\n\n## 方案 3：Wan 2.2 Animate 角色替换\n\n### 工具\n- ComfyUI\n- Wan 2.2 Animate\n- Kijai 节点\n- SageAttention\n\n### 运行平台\n- RunComfy: https://bit.ly/4851Bh9\n- Floyo: https://bit.ly/4oNI0aF\n\n## 选择建议\n- 专业 VFX 流程 → Nuke + WAN 2.1\n- 快速换脸 → LTX 2.3（同种族）\n- 角色风格化替换 → Wan 2.2 Animate",
      "types": ["strategy"],
      "score": 5,
      "source_urls": ["YouTube: Pt7zCjPCyHE", "YouTube: KeAiaFvnnCI", "YouTube: woCP1Q_Htwo"],
      "domain": "AI 视频换脸",
      "tools": ["Nuke", "WAN 2.1", "LTX 2.3", "Wan 2.2", "ComfyUI"],
      "intent": "视频换脸"
    },
    {
      "id": "knowledge-20260409-005",
      "task": "AI Influencer 完整管线：从基础图像到口型同步视频",
      "content": "## 完整管线（Sirio 工作流）\n\n### 工具链\n- KORA Pro（照片级真实感生成）\n- Nano Banana Pro（图像编辑/角色表）\n- Enhancor AI V3（皮肤纹理修复）\n- Enhancor Lip Sync V1/V2（口型同步）\n\n### 步骤详解\n\n| 步骤 | 输入 | 工具 | 输出 | 说明 |\n|------|------|------|------|------|\n| 1 | 提示词 | KORA Pro | 真实基础图像 | 选择强源图像驱动一致性 |\n| 2 | 基础图像 | Nano Banana Pro | 多角度角色表 | 生活方式/工作室场景旋转 |\n| 3 | 角色表 | Enhancor V3 | 修复皮肤纹理 | 解决 AI 塑料皮肤问题 |\n| 4 | 修复图像 + 音频 | Enhancor Lip Sync | 口型同步视频 | 自定义模型 |\n\n### 链接\n- KORA Pro: https://kora.enhancor.ai\n- Enhancor V3: https://enhancor.ai\n- Nano Banana Pro: https://app.enhancor.ai/image-editing\n- Lip Sync: https://app.enhancor.ai/lipsyncing\n\n## Z-Image Turbo 照片级真实感技巧\n\n### 关键发现\n使用以下风格提示词可立即去除\"AI 塑料皮肤\"效果：\n- Phone Photo\n- Selfie\n- Casual Photo\n- Portra Film Photo\n\n### 提示词关键短语\n```\nraw texture, natural imperfections, shallow depth of field,\nslightly imperfect framing, raw photo, unedited look\n```\n\n### 测试提示词示例\n```\nPhotorealistic candid photo of a 22-year-old young Caucasian woman with fair skin,\nnatural light freckles across her nose and cheeks, bright blue eyes,\nlong wavy ash-blonde hair tied in a high messy ponytail,\ndoing glute bridge in modern gym, wearing tight black sports bra and high-waist leggings,\nconfident playful smile looking at camera, sweat glow on skin,\nbright gym lighting with natural window light, energetic and fresh atmosphere,\nvertical portrait, raw mobile photo aesthetic, natural imperfections,\ngym mirror reflection slightly visible, highly realistic,\nsharp focus on face and body, raw texture, natural imperfections,\nshallow depth of field, sharp focus on subject,\nslightly imperfect framing, raw photo, unedited look\n```\n\n## HeyGen 照片转视频口播\n\n### 步骤\n1. OpenArt 生成真实基础图像\n2. HeyGen 选择声音\n3. 编写脚本\n4. 调整运动获得自然表情\n5. 导出工作室级口播视频\n\n### 提示词示例\n```\nFront facing photo of a confident young woman with flowing black hair,\nstriking emerald green eyes with long dark lashes,\ngentle red lips with a soft, genuine smile,\nwearing a casual cream sweater, soft natural lighting,\nprofessional photography style, looking directly at camera, plain background.\n```\n",
      "types": ["case"],
      "score": 5,
      "source_urls": ["YouTube: NpfEKTXBxrY", "YouTube: 1mGQgEnaLfA", "YouTube: cFK_nmLiFys"],
      "domain": "AI Influencer 生成",
      "tools": ["KORA Pro", "Nano Banana Pro", "Enhancor", "HeyGen"],
      "intent": "AI 虚拟人生成"
    },
    {
      "id": "knowledge-20260409-006",
      "task": "提示词工程：去塑料感与多镜头网格生成技巧",
      "content": "## 去塑料感提示词技巧\n\n### 核心关键词（必须添加）\n```\nraw texture（原始纹理）\nnatural imperfections（自然瑕疵）\nshallow depth of field（浅景深）\nslightly imperfect framing（略不完美的构图）\nraw photo（原始照片）\nunedited look（未编辑外观）\n```\n\n### 风格关键词（选择使用）\n- Phone Photo（手机照片）\n- Selfie（自拍）\n- Casual Photo（休闲照片）\n- Portra Film Photo（Portra 胶片照片）\n\n### 效果\n立即去除\"AI 塑料皮肤\"效果，获得真实照片质感\n\n## 多镜头电影网格提示词\n\n### 3×3 网格模板（Tao Prompts）\n```\nA cinematic 3×3 grid presenting multiple camera angles\n{subject and location}\n[主体和场景描述]\nEach frame shows a distinct shot:\n- extreme close-up（极端特写）\n- medium shot（中景）\n- over-the-shoulder（过肩镜头）\n- wide shot（广角）\n- high-angle top-down（高角度俯视）\n- low-angle（低角度）\n- profile（侧面）\n- three-quarter rear view（四分之三后视图）\n- full back view（全后视图）\n{style keywords}\nNatural lighting, muted earth tones, soft fog, minimal color saturation,\nshallow depth of field, cinematic framing.\nConsistent subject appearance across all frames,\nfilm-still aesthetic, professional cinematography reference sheet style,\nultra-realistic photography.\nImages stacked together with no grid lines and no borders. No Text\n```\n\n### 角色表布局关键词\n```\ntwo horizontal rows / four vertical columns\nfront view, left profile, right profile, back view\nfull-body above, portrait below\nA-pose, consistent scale, clean panel separation\n```\n\n## 平台检测绕过技巧（Seedance 2.0）\n\n### 问题\nDreamina/CapCut 上的\"Face detected in your media\"错误\n\n### 解决方案\n使用 Nano Banana 2 将照片级角色表转换为半写实游戏资产\n\n### 转换提示词\n```\nConvert this character sheet into a slightly stylized game character render.\nKeep the face highly recognizable, preserve facial features, bone structure,\nrealistic proportions, and subtle skin detail.\nReduce photorealism slightly.\nUse cleaner shapes, polished game-art rendering, and slightly simplified textures.\nKeep the original composition, pose, outfit, and character design unchanged.\nHigh-quality modern AAA video game character art.\n```\n\n### 工作流\n1. 生成照片级角色表\n2. 使用上述提示词转换为半写实风格\n3. 绕过平台真人检测\n4. 使用 Seedance 2.0 生成视频\n5. 后期可恢复照片级真实感",
      "types": ["strategy"],
      "score": 5,
      "source_urls": ["YouTube: 1mGQgEnaLfA", "YouTube: OA5g63cBp7A", "YouTube: DqWmHvIIcww"],
      "domain": "AI 提示词工程",
      "tools": ["Nano Banana Pro", "Seedance 2.0"],
      "intent": "提示词优化"
    }
  ]
}