Pre-processing routines for Field Latent VAEs and Field Latent Diffusion Models complimenting the official code release
CMake and Eigen are required for installation.
Clone this repository and its submodules
$ git clone --recurse-submodules https://github.com/twmitchel/field_latent_preprocess.git
Afterwards, run ''' pip install -r requirements.txt ''' in the cloned directory.
The script preprocess_FLVAE.py generates train + test splits of planar triangulations for pre-training the FL-VAE on images as described in the paper. The data is stored as .tfrecords files. Set the SAVE_PATH variable on line 20 of the file to the path of the directory where you want to store the data. Afterwards, run python3 preprocess_FLVAE.py to generate the data.
The script preprocess_FLDM.py generates train + test splits for training the diffusion models on a single textured mesh as described in the paper. The data is stored as .tfrecords files. Set the MESH_FILE variable on line 46 to be the location of the .obj mesh file which includes texture uv coordinates per triangle. Set the TEXTURE_FILE variable to the on line 47 to be the location of the .png or .jpg texture file. Set SAVE_PATH on line 48 to the path of the directory where you want to store the data. Afterwards, run python3 preprocess_FLDM.py to generate the data.
Note that these .tfrecords files can be quite large (up to 100 GB) so choose a suitable location accordingly. The file size can be decreased by setting the TARGET_VERTS variable on line 21 to a lower number (say 10K) though this will probably decrease the quality of the generated textures and the FL-VAE should be pre-trained with a larger latent dimension and with coarser triangulations.