planto3d (Voxa)

Upload the plan of a room (JPEG, PNG or PDF) and get an exact, true-scale 3D model of it, shot like an interior photographer would: from the main door at eye level, from behind the bed or the sofa, from the corners and from above. Change anything by typing ("add bed back", "change side table design", "delete that"), turn any shot into a photoreal render with Google's Nano Banana, or take the model to Blender.

How it works

1. Upload a room plan (JPEG / PNG / PDF; draw a box if the drawing has several rooms).
2. Read once. A Gemini vision model reports what is drawn: the room's inner wall lines, the printed dimensions, a box per door, window and furniture symbol. That is the only AI step in building the model.
3. Exact model. Deterministic code turns that reading into a RoomModel in metres: true scale from the printed dimensions, doors and windows on their walls, every piece of furniture at a real size, turned so its back is on its wall and pushed flush against it.
4. Raw shots. The model is built in Three.js with a physically based pipeline (image-based light, a shadow-casting sun through the window, area light at the windows, ambient occlusion, filmic tone mapping), so a raw shot already reads like a semi-rendered Blender frame. The first shot is taken from the main door at eye level (1.55 m). Others: behind the bed or sofa, facing it, the corners and a top view.
5. Edit by typing or in the panels. Commands compile to small, validated operations on the model, with undo and redo. Phrasing the local parser does not know goes to the AI, which returns the same operations.
6. Make realistic (optional) sends a raw shot to Nano Banana (Pro) with instructions to keep every wall, opening and piece exactly where it is and change only materials and light.
7. Download for Blender: room.glb plus render_in_blender.py, which adds every camera, the lighting and Cycles settings.
8. Render in Blender (optional) renders any shot with real Blender Cycles on your own render server.

No AI ever draws or places the room, so there is nothing for it to hallucinate. The model is the single source of truth for the 3D view, every shot, every edit and the Blender export.

AI keys

One key, either kind:

• a Google AI Studio key (AIza…, https://aistudio.google.com/apikey): Gemini reads the plan and handles free-form commands, and Nano Banana makes the realistic renders. The app picks the best models the key can use.
• a kie.ai key (https://kie.ai/api-key): the same jobs through kie.ai's Gemini and nano-banana-pro endpoints.

The 3D model, shots, edits and the Blender export need no key.

Setup and run

npm install
cp .env.local.example .env.local   # optional: a server key for the server build
npm run dev        # http://localhost:3000
npm run build      # production build (also lints and typechecks)
npm run start
npm run lint
npm run typecheck

Deploy

• GitHub Pages (static): .github/workflows/pages.yml builds a static export on every push to main. Visitors paste their own key at the bottom of the page; it stays in their browser's localStorage and goes only to Google or kie.ai.
• Server (Vercel / Node): the default build adds app/api/ai, which uses GOOGLE_API_KEY (or GEMINI_API_KEY, or KIE_API_KEY) for visitors who have not saved a key of their own.

Blender render server

blender-server/ is a small service that renders any shot with real Blender Cycles, from the same model and camera. In the app, every shot card has 🎬 Render in Blender. You host it yourself on a GPU machine, Cloud Run with an L4 GPU, or any Docker host. Then connect it at the bottom of the app with its URL and token. See blender-server/README.md for running, Docker and deployment.

Blender

Unzip the package, then in Blender go to Scripting, open render_in_blender.py and run it. Or render every shot from a terminal:

blender --background --python render_in_blender.py -- --render --samples 128

Cameras that stand behind a wall cut it away with Blender's clip start, the same "clipping camera" the app uses.