Interactive Demo of FloorGenT

FloorGenT is a vector-based generative model of partial floor plans. On the left is a black-and-white canvas, draw a floor plan and let the FloorGenT try to make a floor plan from it. It can be difficult to draw a KTH-like floor plan, so you may want to load a random example from the test set to start from with the button below.

Load random example

Submit your drawing to the network with the button below.

Submit drawing

The star marks the "point of view" location. We also display the model's estimated likelihood (NLL), indicating how likely the generated floor plan is. The output is shown "as is" from the network, no canonicalization or clean-up is done.

The model we use here is using an MLP Mixer for image encoding, going into a Transformer for the sequence model. It uses 793 MiB of GPU memory while sampling. Note that the canvas is larger than the network input, to make it easier to draw. The image is downsampled to 90x90, then padded to 128x128 before it is input to the network.