Plan a room in 2D, walk around it in 3D, and fill it with furniture generated from a single photo. Everything runs locally: the web app works on its own, and an optional generation API drives image-to-3D models such as Hunyuan3D 2.1, TRELLIS.2, TRELLIS, Pixal3D and SAM 3D Objects on your own GPU. You can switch between engines from the app; the API loads and unloads them for you.
Room Setup, itemsmith and glbsmith are three independent tools. Use any one on its own, or connect them:
- itemsmith helps identify products in a room photo from your hints, research reference images and dimensions, and review the matches with you.
- glbsmith uses product photos and width, depth and height to let Codex or Claude Code model the item in Blender and export a GLB.
- Room Setup imports the GLB so you can position it in your own 2D/3D room.
Follow the end-to-end guide for setup, the existing itemsmith → glbsmith handoff, and the manual GLB import into Room Setup. Product matches, dimensions and generated models still need your review; this is not a fully automatic room reconstruction.
Already have a GLB? Use Furniture → Import GLB directly. Room Setup's existing photo-to-3D engines remain a separate option. The companion apps run locally, but their chosen agents use external services.
Found furniture or a gadget you like in an online shop? Take a screenshot of the product photo, enter its size and generate a 3D model. It's placed in your room at real scale, so you can see how it fits and how it looks before you order.
The shop in this video is a mock page made for the demo from our own photo. Generation is shown at 10× speed.
 |
 |
- Floor plan editor: draw walls with grid and corner snapping, Shift to lock angles, drag corners or whole walls (connected walls follow), undo/redo, pan and zoom, rectangle and L-shaped templates, live lengths and floor area.
- Doors and windows: hinged, double, sliding doors and open doorways; sliding, patio, fixed, casement and high windows. Set width, height and sill height, drag them along a wall, flip a door's hinge or swing. They're cut into the 3D walls with frames and glass. Any window can have curtains: choose a fabric and any colour, and draw them back to a tieback.
- Lighting: give a generated lamp a light source that moves with it, or place ceiling lights, pendants, downlights, spots, floor and table lamps, and straight LED tape of any length, set by dragging its ends. (Tape casts no shadows, so its light passes through furniture and walls.) Set brightness (lumens), colour temperature and shadows, move lights in the plan or with 3D arrows, and switch between morning, day, evening and night, with sunlight coming in through the windows.
- 3D room: shadows, environment lighting, a dollhouse cutaway that lowers the walls facing the camera, overview/top/eye-level cameras, and move/rotate gizmos for furniture.
- Walk mode: walk through the room at eye level like a game: WASD or arrow keys to move, the mouse to look around, Shift to run. Walls and tall furniture stop you; doors let you through.
- Finishes: procedural flooring and wallpaper presets, pattern scale, your own photos as textures, and per-wall wallpaper.
- Furniture from a photo: drop or paste an image, enter width × depth × height, pick an engine and generate. Jobs queue up, show their progress, and the result is scaled to your dimensions and placed in the room. Furniture also appears in the plan view, where you can drag it around.
- Shape tweaks: flip models generated upside down or on their side, adjust overall proportions, or pick vertices and deform part of a model.
- Build it with an AI assistant: an MCP server lets Claude (or any MCP client that understands images) recreate your room from photos, a floor plan and product links. Its changes appear live in the app. See docs/mcp.md.
- Projects: save and open everything (walls, finishes, furniture models, thumbnails) as one ZIP. Projects from the original Room Designer app open too.
- Light and dark themes, Japanese and English UI, keyboard shortcuts.
Requirements: Node.js 22.12+, a WebGL-capable desktop browser. The generation API needs Python 3.10+.
npm ci
npm run dev # http://127.0.0.1:5173Planning, finishes and opening saved projects work without the API. To try the generation flow without a GPU, run the API with the built-in mock engine (it returns a box):
python3 -m venv .venv
.venv/bin/pip install -r server/requirements.txt
ROOM_SETUP_MOCK_ENGINE=1 npm run api # in a second terminal
| Engine | Output | Notes |
|---|---|---|
| Hunyuan3D 2.1 | Shape, optional PBR texture | Texturing is slower and needs more GPU memory. Has its own, more restrictive license (see below). |
| TRELLIS.2 | Shape with PBR materials | Most detailed results. |
| TRELLIS | Coloured mesh | Fast. |
| Pixal3D | Shape with PBR materials | Follows the photo closely. MIT, but downloads Meta's DINOv3 (own license). |
| SAM 3D Objects | Textured mesh | Robust on cluttered real-world photos. Meta SAM License; gated weights. |
All engines need an NVIDIA GPU with CUDA and have been tested on Linux with an RTX 4090 (24 GB).
| Engine | Minimum VRAM | Recommended |
|---|---|---|
| Hunyuan3D 2.1 | 10 GB (shape only) | 24 GB (with texture; texturing alone needs about 21 GB) |
| TRELLIS.2 | 24 GB | 24 GB+ |
| TRELLIS | 16 GB | 16 GB+ |
| Pixal3D | 24 GB (low-VRAM mode, the default here) | 24 GB+ |
| SAM 3D Objects | 24 GB with this project's settings (upstream asks for 32 GB) | 32 GB |
Minimums come from each upstream project except where noted; only one engine uses the GPU at a time, so the largest one you install sets the requirement.
Each engine stays in its own Python environment; this project does not bundle any model code or weights. Install the ones you want by following their upstream instructions, then point the API at them:
cp .env.example .env # set HUNYUAN3D_PYTHON / HUNYUAN3D_ROOT, TRELLIS2_…, TRELLIS_…
npm run apiThe engine list in the app shows which engines are set up and whether each is idle, loading or ready. The selected engine is loaded when a job needs it; whichever engine held the GPU before is shut down first, so only one model occupies VRAM at a time. See docs/engines.md for setup details, GPU memory, troubleshooting, and how to add another engine.
With the API and the app running, register the MCP server with your assistant:
claude mcp add room-setup -- "$PWD/.venv/bin/python" "$PWD/server/mcp_server.py"Then send it your room photos (from the middle of the room towards each corner), the floor plan and links to your furniture, and ask it to build the room. Setup for Claude Desktop, tips and the tool list are in docs/mcp.md.
- Room tab: pick a template or draw walls in the plan (
W). UseVto move corners and walls,Eto erase. Set the ceiling height. Under Doors & windows, pick a preset and click a wall to place it; select it to change its size, sill height or swing. - Finishes tab: choose flooring and wallpaper, adjust the pattern size, or use “+” for your own image. To change a single wall, select it in the plan and choose “Selected wall only”.
- Furniture tab: drop/paste a photo, enter its size, pick an engine and click Generate 3D model. You can queue several.
- Switch to 3D to look around, click furniture to select it, and drag the gizmo. The inspector lets you rename, position, rotate, duplicate, delete or adjust the shape.
- Save downloads the project as a ZIP (
room-project.zipuntil you name it), and Save as in the arrow menu next to it lets you choose the name. Open restores it. Later saves keep the name of the file you saved or opened. While the API is running, the current room is also kept by the API and comes back after a reload. Without the API, nothing is saved automatically.
| Shortcut | Action |
|---|---|
W / V / E |
Draw walls / select & move / erase (plan) |
Shift while drawing |
Lock to 45° steps |
Space + drag, middle or right drag, wheel |
Pan and zoom the plan |
F |
Fit the plan to the view |
Ctrl+Z / Ctrl+Shift+Z |
Undo / redo: walls, doors and windows, finishes, lights and furniture, also with the buttons in both views |
R / Shift+R |
Rotate the selected furniture by 90° |
Delete |
Delete the selected furniture or wall |
Ctrl+S |
Save the project |
Ctrl+Shift+S |
Save as |
Esc |
Cancel / deselect |
flowchart LR
B[Browser app<br/>Vite · Three.js] -- /api, /generated<br/>(Vite proxy) --> G[Gateway<br/>FastAPI]
G -- JSON lines --> H[Hunyuan3D worker<br/>own Python env]
G -. one at a time .-> T2[TRELLIS.2 worker]
G -.-> T1[TRELLIS worker]
The browser only talks to its own origin; Vite forwards API calls, so the gateway stays on 127.0.0.1 and needs no CORS. Developer notes and the API reference are in docs/development.md.
Photos you feed into the generator, such as product images from online shops, usually belong to someone else. Only use images you have the right to use. Please keep models generated from other people's images to personal use (planning your own room): don't redistribute or sell them, and take care with logos and trademarks that show up in textures. The best input is a photo you took yourself.
This is a single-user tool for your own machine. The app and API have no authentication (the API does refuse requests from other websites). Keep both bound to localhost (the default). If you expose the dev server on a LAN with HOST=0.0.0.0, do so only on a network you trust, and never on the public Internet.
Room Setup's own code is MIT. Generation engines, model weights and their outputs are covered by their own terms. In particular, Tencent Hunyuan 3D 2.1 has territorial restrictions and other conditions that this project's MIT license does not change. Read THIRD_PARTY_NOTICES.md before using or distributing generated models.
Contributions are welcome; see CONTRIBUTING.md.