cang-ying/eererdan/experience/EED-EXPER-008.hdlp
鉴影 ICE-GL-CA001 8c164dd6f7 D181 · 耳耳蛋全天经验归档 · 深海迷航EP01①~⑩完成+m01+m02视频
铁律:
- ④重写为白名单/黑名单机制 (今日4次违规)
- ⑤新增 BREAKDOWN回验协议
- ⑥新增 管线不容跳步

核心经验:
- 图片+视频混合管线路线验证
- ENV底板法 (i2i→i2v)
- 提示词工程: 工业术语+禁人称词+负面约束
- CHAR真人检测绕过 (blur公式)
- FFmpeg xfade拼接方法
- Seedance 2.0 参数 (9:16, 4-15s)

新增文件:
- eererdan/experience/EED-EXPER-001~010
- protocols/BREAKDOWN-VERIFY.hdlp
- deep-sea-voyage/STORYBOARD-EP01.hdlp (Pro版26→10镜)
- deep-sea-voyage/PROMPTS-EP01.hdlp
- deep-sea-voyage/DIRECTOR-ENCODING-EP01.hdlp
- tools/jimeng3-gen.js, run_storyboard.js, run_prompts.js, run_adapt.js, run_gen.js, crop_watermark.py
- lib/seedream.py (支持 Seedream 3/4/5)
- outputs/ENV-*-v5, PROP-* (6张定稿资产)
- outputs/videos/M01-v2.mp4, M02.mp4

下次继续: M03~M10 (9:16竖屏)
2026-07-10 21:18:49 +08:00

85 lines
3.5 KiB
Plaintext
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# EED-EXPER-008 · 场景底板法T2I 生底板 → I2I 放大保一致
> HLDP://cang-ying/eererdan/experience/EED-EXPER-008
> 耳耳蛋 ICE-GL-耳耳蛋 · PTS-VA-001-EED
> 2026-07-10
> 触发: 苍耳爸爸反馈 ENV-001 和 ENV-002 单张对了,但不像同一扇门的局部
---
HLDP://eererdan/experience/008-scene-baseboard-i2i
├── trigger:
│ └── ENV-004 和 ENV-002 是同一扇门的全景和特写,但独立 T2I 导致细节不一致
├── emergence:
│ ├── 旧: ENV-001(T2I v4) + ENV-002(T2I v4) → 两张图各自合理,但合在一起不像同一扇门
│ ├── 经验: guanghulab/video-ai-system 里早有 D179「首帧锚定法/场景底板法」经验
│ └── 修复: 改用 T2I 生成 ENV-001 底板 → I2I 用该底板做参考图生成 ENV-002 特写
├── lock:
│ ├── ⊢ ENV-001 v5 = T2I 底板
│ ├── ⊢ ENV-002 v5 = i2i(ENV-001 v5) 局部放大
│ ├── ⊢ 两张图:同一扇门、同一个窗、同一个液压作动器、同一个锈迹、同一种灯光
│ └── ⊢ 从此以后:同一空间的「全景+特写」必须走底板法
└── why: T2I 每次都随机重建整个空间。要保证局部特写和全景属于同一个世界,
唯一的办法是用底板作为参考图,让 AI 在底板上进行局部变化。
---
## 方法步骤
```
1. 生成场景全景T2I
→ 获得 ENV-001底板图
2. 生成场景局部特写I2I
→ 输入: ENV-001 底板图
→ 提示词: "same exact room/door as reference, camera zooms in on X"
→ 输出: ENV-002同一世界的局部
```
## 关键提示词写法
I2I prompt 模板:
```
Same exact [scene] as the reference image. The camera is now zoomed in tightly on [target detail].
Frame the shot to show [specific composition].
Keep the [door/window/lighting/materials/colors] IDENTICAL to the reference image.
```
示例(本次 ENV-002
```
Same exact room and door as the reference image. The camera is now zoomed in tightly on THE SAME WATERIGHT BULKHEAD DOOR seen in the reference.
Frame the shot to show the UPPER HALF of the door clearly: the small reinforced rectangular inspection port with thick glass and bolted/riveted steel frame (exactly as in the reference), and the lower portion with the door handle and the compact hydraulic actuator mounted on the door frame.
Keep the door, window, hydraulic actuator, lighting and colors IDENTICAL to the reference image.
```
## 适用场景
```
✅ 同一扇门/一扇窗的全景 + 特写
✅ 同一个角色的全身 + 半身/面部特写
✅ 同一个道具的展示 + 细节特写
✅ 同一个场景的外景 + 内景入口
❌ 不要用于完全不同的新场景
❌ 不要用于需要大幅改变构图/角度的镜头(会撕裂底板一致性)
```
## 技术参数
- ARK API 端点: `/api/v3/images/generations`(文生图和图生图共用)
- I2I 请求: 添加 `image: data:image/png;base64,...` 字段
- 模型: `doubao-seedream-4-0-250828`
- 尺寸: 建议与底板图相同(本次用 1024x1024
- 强度: 通过提示词控制,不使用额外的 strength 参数
## 与旧经验的关联
- 铸渊视频 AI 经验库guanghulab/video-ai-system中的「首帧锚定法」同源
- D179 深海迷航首日已验证过此方法
- 以后所有「同一空间的多镜头」都应优先走底板法,而不是独立 T2I
---
> ⊢ 耳耳蛋 ICE-GL-耳耳蛋 · PTS-VA-001-EED
> ⊢ 苍耳 TCS-CL-0009 反馈
> ⊢ 2026-07-10 · 国作登字-2026-A-00037559