ArchVoxel Repo

The model should create voxel representations of buildings using latent diffusion and an intermediate VAE.

The training procedure is as follows:

Use Vae.py to train the VAE to compress 32³ voxel chunks into a N-dimensional vector. The buildings in /data have a varying size and are split into these chunks. Top row shows reference, bottom row is the reconstruction.
Latent_preprocess.py takes the whole building and calculates a tensor of VAE-compressed representations. This could be done during the diffusion model training, but it's quite slow and can be preprocessed.
traindiffusion.py is then used to train the diffusion model based on the latent representation. Intermediate results are saved in the binvox format.
interpolate.py can be used to visualize the latent space interpolation.

TODOs

Enlarge train data set.
The diffusion model has only been trained for 100k steps with a small model. Still figuring out the best model complexity and training scheduling!
Eventually increase voxel resolution and add more data into the voxel (not just binary visibility) to reconstruct geometry