Four to six regular photos of the same room, or just a text description, become a photorealistic equirectangular panorama you can rotate like a virtual tour. No 360 camera, scan, or 3D software needed.

The AI understands how the photos connect (the same sofa in two photos, the same window from two angles) and reconstructs the room as a single space in a 2:1 format at 4096×2048, ready for any 360° viewer.
Preserves walls, doors, windows, furniture, and materials. Doesn't redesign or invent objects; each piece of furniture appears only once. What no photo shows is inferred conservatively.
No photos? Describe the room in text. AI creates five consistent photos of the same space and then assembles the 360° panorama. Useful for concepts, previews, and content.
Exact 2:1 ratio, minimum resolution of 4096×2048, score seam edges, automatic regeneration if the seam fails, and final smoothing. Everything is included in the report.
It's an AI reconstruction, not a measurement. It's suitable for real estate marketing, previews, virtual tours, and input for AI video. It doesn't replace a LiDAR scan.
Gemini reads all the photos as a single space: walls, doors, windows, furniture, lighting, the estimated direction of each photo, and what's missing.
The photos are used as references for a model with native 2:1 support. The default is Flux 2 Klein via Freepik/Magnific; GPT Image 2 and Gemini via OpenRouter are alternatives.
Corrected ratio, resampling up to 4096×2048, seam score (≤12 good, ≤30 acceptable, >30 visible), one automatic regeneration, and crossfade at the edges.
Keys are read at runtime from your environment file. Nothing is copied into the project.
FastAPI, Pillow, NumPy, and httpx.
# in the project folder pip install -r requirements.txt
Freepik/Magnific for generation, Gemini for vision. OpenRouter is optional.
# ~/projetos/wifi/.env MAGNIFIC_API_KEY=... GEMINI_API_KEY=... OPENROUTER_API_KEY=... # optional
4 to 6 JPG/PNG photos of the same room from different angles. Good coverage is better than five nearly identical photos.
sala/ ├── 01.jpg # turn toward the sofa ├── 02.jpg # turn toward the window └── 03.jpg # turn toward the entrance
They all use the same engine. The output goes to ~/projetos/output/fotos3d/<nome>/ with the PNG, analysis, prompt, and report.
Choose a folder and give it a name. The default model is Flux 2 Klein (10 Magnific credits per image).
cd ~/projetos/fotos3d python3 -m engine.pipeline /caminho/fotos-da-sala --nome sala-302 # ~1 min
Describe the room. The AI creates five consistent photos and follows the same workflow.
python3 -m engine.pipeline --assunto "sala moderna, sofá cinza em L, janela grande com vista pro jardim, piso de madeira clara" --nome demo
Switch models whenever you want more references or native 4K. The report records which one was used.
python3 -m engine.pipeline sala/ --nome sala-302 --modelo gpt-image-2 --size 4K python3 -m engine.pipeline sala/ --nome sala-302 --modelo or-gemini-pro # via OpenRouter
Drag in the photos (or paste the folder path), choose a model with its cost shown beside it, track progress, and rotate the result in the built-in viewer. A gallery shows everything generated so far.
uvicorn web.app:app --port 8360 # open http://127.0.0.1:8360
Always check in the viewer, not in the flat image: duplicated furniture, repeated doors, impossible walls, lighting jumps, or breaks at the turning point. If something is wrong, regenerate or switch models.
# standalone viewer: open it in your browser and drag in the PNG viewer/index.html # report with aspect ratio, resolution, and stitching score ~/projetos/output/fotos3d/sala-302/relatorio.json
Install once, then ask in natural language: "use the property-360 skill on the photos in folder X".
ln -s ~/projetos/fotos3d/skill/property-360 ~/.claude/skills/property-360
| Model | Provider | native 2:1 | Test result | Cost |
|---|---|---|---|---|
flux-klein (default) | Freepik / Magnific | yes, resampled to 2048×1024 | true equirectangular, highly faithful | 10 credits |
gpt-image-2 | Freepik / Magnific | yes, up to 4K, 16 refs | best quality: correct spherical projection, good stitching | ~45 to 90 credits |
or-gemini-pro | OpenRouter | no (resampled from 16:9) | Plausible 360°, less faithful to details | ~US$ 0,13 |
or-gemini-flash | OpenRouter | no | tends to come out wide-angle, not 360 | ~US$ 0,04 |
Above are the input photos. Below is the generated panorama (it looks stretched on a flat screen, but wraps around 360° in the viewer).





Version 1.0 handles one room at a time. Future phases will connect the rooms into a tour.