PyTorch implementation of the solid texture synthesis model in the journal article On Demand Solid Texture Synthesis Using Deep 3D Networks published in Computer Graphics Forum. DOI:10.1111/cgf.13889
Code based on Gatys' Neural Style Transfer implementation
The python script train_slice.py trains a generator network and produces the file params.pytorch that contains the trained parameters.
It requires the libraries: PIL and PyTorch.
The VGG-19 perceptual loss between 2D images uses Gatys' implementation mentioned above.
To run the code you need to get the pytorch VGG19-Model from the bethge lab using the script download_models.sh from https://github.com/leongatys/PytorchNeuralStyleTransfer
Using display is optional.
The names of the example textures and the associated training directions are defined by the lists input_names and directions.
The example textures go in the Textures folder.
The python script sample_on_demand.py loads the trained parameters and synthesizes a block of texture of sizes total_H, total_W, total_D formed of blocks of sizes piece_height, piece_width, piece_depth.
It requires the libraries: PIL, PyTorch and cupy
Indicate the location of the trained model in model_folder.
The output file *.npy is a numpy 4D array with the number of channels (BGR) in the first dimension and the spatial dimensions in the next three dimensions.
One visualization option is ParaView. It requires to convert the *.npy file to VTK. We used the library PyEVTK (see gridToVTK).
Additionally we provide a PyOpenGL script render_cube.py that loads the *.npy file and uses it to render a simple cube.
The code requires OpenGL and the python libraries: pyopengl, sdl2 and pyglm
texture defines the path to the texture (*.npy) to apply. Use the key 's' to stop and start the rotation of cube