做 AI 漫剧、动态漫、解说漫、短剧、分镜或影视化的场景资产时，最头疼的是同一个场景换个机位就"穿帮"：左右物体对不上、生图模型很难根据图片场景想象出反打画面，导致创作者卡在场景资产创作无法推进下一步。
本项目把这件事变成一条几何可追溯的闭环流水线：
- 阶段 A0（冷启动推断）：你只发一张图、没多说信息时，先根据画面反推“这是什么场所的哪个功能分区、反打/左/右各应该是什么区”（例如一眼认出是咖啡店吧台制作区，反打应是顾客点单/座位区），先向你陈述并等你确认，不擅自开画；
- 阶段 A（成图）：分析场景图 → 抽象成结构化平面 JSON → 用脚本渲染出一张标准黑白俯视 CAD 工程图（不是让 AI 乱画示意图，方位严格可靠）；
- 阶段 B（出视角）：你在平面图上画箭头（箭头尖＝想看到的那面），或直接说“反打 / 左侧 / 右侧 / 转 X 度” → 几何推演出新机位与平面布局 → 以原图为参考生成同一空间、同一光影材质的新视角图，可反复指定；
- 阶段 C（四宫格总检）：所需视角出齐后，脚本自动拼一张固定机位顺序的 2×2 四宫格，统一核对位置关系有没有错乱、风格色调是否一致，有问题就改该视角再重拼。
它是一个标准 Agent Skill（
SKILL.md+scripts/+references/），遵循开放的 Agent Skills 目录规范，可被豆包、Claude（Claude Code / Desktop）、兼容该规范的客户端直接加载；脚本本身是纯 Python，也可单独命令行使用。
下列示例图均由本流程实际产出（AI 生成，仅作功能演示）。
| 场景原图 Original | 俯视 CAD 平面图 Floor plan |
|---|---|
 |
 |
| 在图上画箭头指定方向 Pick a view | 生成的反打视角 Reverse view |
 |
 |
长条空间最容易出错——垂直看墙时被看墙应横向平展，只有沿走廊长轴看才有两侧向灭点汇聚。下图严格区分了这两种透视：
四机位闭环（原视角 / 反打 / 左侧 / 右侧），注意反打时挂件左右正确互换：
U 形院落反打时，三栋楼全部转到身后、不入镜，画面主体是机位背后的医院大门——这类围合空间的"谁在前方 / 谁在身后"判定也已沉淀进规则：
所有视角出齐后，make_grid.py 按固定机位顺序（左上原视角 / 右上反打 / 左下左 / 右下右）拼一张四宫格，一眼看清位置关系与风格是否统一。下图中四格冷蓝色调一致、对角的原视角与反打左右正确互换、左右两格互为反向纵深，即判定自洽：
flowchart TD
A[场景原图] --> A0{信息是否充足?}
A0 -->|只给图/没说场景| P[阶段A0 反推场所与各方向分区<br/>向用户陈述推断]
P --> Q{用户确认?}
Q -->|纠正| P
Q -->|确认/本就充足| B[视觉分析<br/>机位/墙体/门窗/家具]
A0 -->|已说清布局| B
B --> C[结构化 plan JSON]
C --> D[floorplan.py 渲染<br/>黑白俯视 CAD]
D --> E{用户指定方向}
E -->|画箭头 / 反打 / 左 / 右| F[几何推演新机位<br/>左右互换·可见性·纵深]
F --> G[以原图为参考<br/>图生图]
G --> H[对照平面图逐张回查]
H -->|不符则改几何重生成| G
H -->|本视角通过| K{所需视角出齐?}
K -->|否,继续指定| E
K -->|是| L[阶段C make_grid.py<br/>拼 2x2 四宫格总检]
L --> M{位置/风格/数量自洽?}
M -->|某格错乱| F
M -->|全部通过| I[整体交付四宫格+各视角]
几条不可违反的几何铁律（也是本项目的核心价值）：
- 箭头尖指向 ＝ 相机新朝向 ＝ 画面正前方看到的那面；箭头是视线方向，不是相机移动方向。
- 反打必做左右互换：平面点绕场景中心旋转 180°，
(x,y) → (2Cx−x, 2Cy−y)，原左变新右、原右变新左。 - 长条空间先判透视类型：视线垂直长墙 → 被看墙横向平展、无侧墙汇聚；视线沿长轴 → 两侧墙才向灭点汇聚。
- 围合空间先判前后：落在新机位后向半球的建筑一律不画，避免"反打还看到原尽头主楼"。
- 画面外信息不编造：机位背后拍不到的区域只能来自用户补充或原图可合理延伸，缺失先问。
需要 Python 3.9+：floorplan.py 依赖 matplotlib（会自动带上 numpy），make_grid.py 依赖 Pillow：
git clone https://github.com/EdwinZDZ/scene-view-planner.git
cd scene-view-planner
pip install -r requirements.txt脚本会自动挑选系统中文字体（Windows / macOS / Linux 均已做回退），中文标注不会变方块。
本目录本身就是一个合法 Skill：
- 豆包 / Claude Code / Claude Desktop 或兼容 Agent Skills 规范的客户端：把整个
scene-view-planner/目录放进其 skills 目录（如 Claude Code 的~/.claude/skills/），随后直接对 Agent 发场景图即可，它会按SKILL.md自动走完 A0→A→B→C 完整闭环。 - ChatGPT（自定义 GPT）：把
SKILL.md与references/内容贴入 Instructions / Knowledge，并把scripts/floorplan.py上传到 Knowledge，勾选 Code Interpreter。详细迁移步骤见 SKILL.md 顶部说明与references/。
- 参照下方 plan JSON 速查 或
references/cad_symbols.md，把空间写成plan.json； - 运行：
# 生成黑白俯视 PNG（可同时导出 SVG 矢量图）
python scripts/floorplan.py plan.json -o floorplan.png --svg floorplan.svg
# 用户确定方向后，叠加一个红色箭头做机位自检（x,y,方向）
python scripts/floorplan.py plan.json -o marked.png --arrow 3,3.4,S- 把
floorplan.png交给任意图生图模型，按references/camera_geometry.md第 6 节的五段提示词模板生成新视角。
当你只发图、没说明这是什么地方时，Agent 不会直接画平面图，而是先按 references/scene_inference.md 做三步反推并向你确认：
- 识别场所与当前分区：从设备家具、标识、材质、采光取证（咖啡机+磨豆机+菜单黑板 → 咖啡店吧台制作区）；
- 推断四方向分区：以镜头朝向为前，预判正前方 / 反打 / 左侧 / 右侧 / 画面外各是什么区（如咖啡店反打应是点单取餐与堂食座位区）；
- 陈述并等你确认：明确区分"画面看到的事实"与"经验推断"，你确认或纠正后才进入阶段 A。这样画面外布局在动手前就对齐，避免反打凭空编造。
- 读图：定原机位位置与视线方向、空间轮廓与长宽比、各面墙上的门窗与固定物、主要家具沿墙位置；
- 补画面外信息：用户说"对面是墙还有一道门""背后是大门"等务必纳入，没说又影响目标视角时先问；
- 写 plan JSON：北朝上、x 东 y 北，只要求相对比例与拓扑正确，不要求真实尺寸；
- 渲染并回读：必须实际打开输出 PNG，对照原图核对机位、门所在墙、左右排布，不对就改 JSON 重渲染。
口头方向换算（原朝向 F0，北朝上）：
| 原朝向 F0 | 反打 180° | 左侧视角 | 右侧视角 |
|---|---|---|---|
| N | S | E | W |
| E | W | S | N |
| S | N | W | E |
| W | E | N | S |
生成前先在平面图上"站到新机位"列清：正中尽头是什么、新左右各是什么、谁进出画面、纵深与遮挡顺序；随后用五段模板组织提示词（空间风格锚点 → 新机位 → 画面布局 → 反打/侧视几何关系 → 一致性约束），以原始场景图为首要参考，只转机位、不重新装修。完整模板与实例见 references/camera_geometry.md。
反打 / 左 / 右都生成完，运行（纯 Pillow）：
python scripts/make_grid.py --front 原图.png --reverse 反打.png \
--left 左侧.png --right 右侧.png -o grid.png固定排布为「左上原视角 · 右上反打 · 左下左 · 右下右」，随后按三项清单核对：① 位置关系（相邻 90° 边缘衔接、对角 180° 左右互换且不见同一尽头、左右两格互为反向纵深）；② 风格统一（色温/材质/光照/视高焦段一致）；③ 数量一致（门窗固定物跨视角对得上）。某格出错就回阶段 B 改该视角并重拼，直到四格自洽。详见 references/camera_geometry.md 第 8 节。
openings.kind：door / slide / opening / window / double；wall：N/S/E/W。elements[].type：wall、rect、circle、label、toilet、sink、stall_row、sink_row、counter、arrow、camera。- 未内置的物体一律用
rect / circle / wall组合成俯视外轮廓近似。完整字段、参数与更多示例见references/cad_symbols.md。
scene-view-planner/
├── SKILL.md # Agent 主指令：A0/A/B/C 闭环 + 铁律
├── scripts/
│ ├── floorplan.py # plan JSON → 黑白俯视 CAD（PNG/SVG/红箭头）
│ └── make_grid.py # 多视角 → 固定顺序 2×2 四宫格（阶段C自检）
├── references/
│ ├── scene_inference.md # 阶段A0 场景反推三步法 + 空间先验库 + 确认模板
│ ├── cad_symbols.md # 透视图反推平面 + 全部图块字典
│ └── camera_geometry.md # 机位几何推演 + 提示词模板 + 四宫格自检清单
├── examples/ # 三个真实案例（原图/平面/箭头/各视角）
│ ├── bathroom/
│ ├── corridor312/
│ └── courtyard/
├── docs/images/ # README 用的流程、对比与四宫格展示图
├── requirements.txt
└── LICENSE
欢迎提交 Issue（错误机位案例尤其欢迎，这是规则迭代的最佳来源）与 PR：新增图块类型、补充 references/ 里的空间类型判据、增加示例、完善英文版本文档都非常欢迎。提交前请确保 python scripts/floorplan.py 能正常渲染、且不破坏既有示例。
MIT。示例图像由 AI 生成，仅用于功能演示。
Turn one scene photo into a standard top-down CAD floor plan, then render the reverse / left / right / any-angle view of the same space on demand.
When building AI comics, motion comics, short films or storyboards, the hardest part is keeping a space consistent across camera moves — objects swap sides incorrectly, a reverse shot shows the wrong wall, or a corridor turns into a weird "bulge". This project turns it into a geometrically traceable closed loop:
- Stage A0 (cold-start inference) — when you send only one photo with no extra info, the agent first infers the venue and current functional zone (e.g. a café's bar/prep area) and predicts what the reverse/left/right directions should contain (customer counter, seating area…), then states this guess and waits for your confirmation before drawing anything.
- Stage A — analyze the scene photo, abstract it into a structured plan JSON, and render a clean black-and-white top-down CAD plan with a script (not an AI sketch — the orientation is reliable).
- Stage B — you draw an arrow on the plan (arrow tip = the side you want to see) or say "reverse / left / right / rotate X°"; the agent geometrically derives the new camera and generates a new view of the same space with identical lighting and materials, using your original photo as reference. Repeatable for any direction.
- Stage C (grid self-check) — once all needed views exist, a script assembles a fixed-layout 2×2 grid to verify camera relationships and visual consistency at a glance; any faulty cell is regenerated and the grid rebuilt.
One floor plan can be reused for as many camera directions as you need.
It ships as a standard Agent Skill (SKILL.md + scripts/ + references/) following the open Agent Skills folder convention, while the renderer is plain Python you can also run standalone.
Hospital corridor — four camera setups of one "=" shaped space. A view perpendicular to the long wall stays flat and horizontal; only a view along the axis converges to a vanishing point.
U-shaped courtyard — a reverse shot faces the main gate behind the camera. All three buildings stay behind the new camera and leave the frame; the gate becomes the subject.
More per-step images are under examples/ (bathroom/, corridor312/, courtyard/).
- Arrow-tip direction = new camera facing = what is seen straight ahead. It is a viewing direction, not a movement.
- Reverse shot always swaps left/right: rotate 180° around the center,
(x,y) → (2Cx−x, 2Cy−y). - Long & narrow spaces: perpendicular-to-wall views stay flat with no converging side walls; along-the-axis views converge.
- Enclosed (U/L-shaped) spaces: decide per object whether it is in front of or behind the new camera; anything in the rear hemisphere is dropped.
- Never invent off-camera layout — use only user-supplied info or reasonable extension of the original; when only a photo is given, Stage A0 infers the zones and asks for confirmation first.
- Cold-start gate: with no venue/context supplied, infer and confirm the scene type and off-screen zones before rendering the plan.
- Final 2×2 grid self-check: after all views are done, assemble them (front / reverse / left / right) and re-verify geometry, style and object counts as a whole.
git clone https://github.com/EdwinZDZ/scene-view-planner.git
cd scene-view-planner
pip install -r requirements.txt # Python 3.9+, matplotlib(+numpy) for CAD, Pillow for the gridA CJK font is auto-selected on Windows / macOS / Linux, so Chinese labels render correctly.
As an Agent Skill (recommended): drop the whole folder into your runtime's skills directory (e.g. ~/.claude/skills/ for Claude Code); compatible agents then follow SKILL.md automatically. For a ChatGPT custom GPT, paste SKILL.md + references/ into Instructions/Knowledge, upload scripts/floorplan.py to Knowledge, and enable Code Interpreter.
Standalone rendering:
python scripts/floorplan.py plan.json -o floorplan.png --svg floorplan.svg
python scripts/floorplan.py plan.json -o marked.png --arrow 3,3.4,S # overlay a red arrow
# Stage C: stitch all views into a fixed-layout 2x2 self-check grid
python scripts/make_grid.py --front f.png --reverse r.png --left l.png --right r2.png -o grid.pngWith only a photo and no context, Stage A0 first infers the venue/zones and asks you to confirm before the plan is drawn (see references/scene_inference.md). After every view is generated, Stage C rebuilds the 2×2 grid and checks camera geometry, style and object counts together (see section 8 of references/camera_geometry.md).
Verbal direction map (original facing F0):
| F0 | Reverse 180° | Left | Right |
|---|---|---|---|
| N | S | E | W |
| E | W | S | N |
| S | N | W | E |
| W | E | N | S |
See references/cad_symbols.md for the full plan-JSON schema/block catalog and references/camera_geometry.md for visibility reasoning and the five-part image-to-image prompt template.
scene-view-planner/
├── SKILL.md # main agent instructions (A0/A/B/C loop + rules)
├── scripts/
│ ├── floorplan.py # plan JSON -> top-down CAD (PNG/SVG/arrow)
│ └── make_grid.py # views -> fixed 2x2 self-check grid (Stage C)
├── references/
│ ├── scene_inference.md # Stage A0 scene inference + venue prior library
│ ├── cad_symbols.md # block catalog / plan-JSON schema
│ └── camera_geometry.md # camera reasoning + prompt template + grid checklist
├── examples/ # three real worked cases
└── docs/images/ # diagrams used by this README
Issues (especially wrong-camera cases) and PRs are welcome — new block types, more space-type rules in references/, new examples, and English docs. Released under the MIT License. Example images are AI-generated and used for demonstration only.
{ "style": {"wall": 3.0, "inner": 2.0, "item": 1.8, "compass": true}, "room": {"bounds": {"xmin": 0, "ymin": 0, "xmax": 6, "ymax": 8}}, // 异形用 "outline":[[x,y]...] "openings": [ {"kind": "door", "wall": "S", "at": 3.0, "width": 1.1, "hinge": "start"} ], "elements": [ {"type": "stall_row", "x": 1.8, "y": 8.0, "n": 3, "cell_w": 0.9, "depth": 1.5, "orient": "S", "toilet": true}, {"type": "sink_row", "wall": "E", "x": 5.55, "y0": 2.1, "n": 4, "pitch": 0.85, "w": 0.55, "d": 0.62} ], "camera": {"x": 3.0, "y": 0.7, "facing": "N", "label": "原机位"} }