Floorplans that dream of being furnished apartments

What it does

Plan2Scene takes a floorplan and a set of interior photos, then builds a textured 3D mesh of the residence. When photos don’t cover every surface—which is always—it uses a GNN-based texture propagation network to fill in the gaps. The output is a tileable, seam-corrected 3D scene you can preview in a browser.

The interesting bit

The system doesn’t just slap photos onto walls. It ranks photo crops by “textureness” using a VGG-based scorer, then propagates patterns to unobserved surfaces through a graph neural network trained to guess what the unseen parts should look like. There’s also a dedicated seam-correction step to make everything tileable—because nothing ruins immersion like a repeating wallpaper seam.

Key highlights

• Built around the Rent3D++ dataset, with pretrained models available
• Simulates photo “un-observations” (drop ratios up to 60%) to stress-test texture hallucination
• Optional raster-to-vector floorplan preprocessing via FloorplanTransformation
• Outputs standard scene.json files viewable through SmartScenesToolkit
• Google Colab demo available for quick experimentation

Caveats

• Heavy dependency stack: conda, Embark Studios texture-synthesis CLI, SmartScenesToolkit build from source
• README is essentially a long shell script with prose framing; expect significant manual path configuration
• Custom data pipeline requires converting floorplans to a specific vector annotation format first

Verdict

Worth a look for researchers in indoor scene synthesis or anyone building automated property visualization tools. Skip it if you need a drop-in API or clean Python package—this is research code that wears its complexity openly.