xmaster6y/wgenpatex

Wasserstein Generative Models for Patch-based Texture Synthesis

wgenpatex

Goal of this fork

This fork implements the algorithm 1 bis from the paper https://arxiv.org/abs/2007.03408. This algorithm consists in recovering lost data of an image. To illustrate it, I followed the example of the article by chosing to do inpaiting on an image (demo_texture_1_patched.png).

In addition to the previous file a presentation (in french) highlighting the key points of the article is provided.

Edited from original Readme

Wasserstein Generative Models for Patch-based Texture Synthesis

Code related to the paper https://arxiv.org/abs/2007.03408

• wgenpatex.py contains all the functions
• model.py contains the generative model architecture
• syntax python run_optim_synthesis.py target_image_path --options for running texture synthesis with image optimisation (Alg.1 from paper)
• syntax python run_cnn_synthesis.py target_image_path --options for learning an convolutional neural network texture generator (Alg.2 from paper) (GPU recommanded)
• syntax python run_optim_patching.py target_image_path --options for running image inpaiting with image optimisation (Alg.1bis from paper)

replace --options with any:

-w or --patch_size patch size -nmax or --n_iter_max max iterations of the algorithm -npsi or --n_iter_psi max iterations for psi -nin or --n_patches_in number of patches of the synthetized texture used at each iteration, -1 corresponds to all patches -nout or --n_patches_out number of patches of the target texture used at each iteration, -1 corresponds to all patches -sc or --scales number of scales used

--visu plot intermediate results --save save intermediate results in /tmp folder --keops' use keops package (speed-up)