/optimal-textures

Optimal deep texture generation and style transfer based on Eric Risser's paper

Primary LanguagePythonMIT LicenseMIT

Optimal textures

An implementation of the texture generation algorithm proposed in Optimal Textures: Fast and Robust Texture Synthesis and Style Transfer through Optimal Transport.

Example

Texture generation

python run.py scale.jpg decoder_states/scale/ -n 5

output

Style transfer

python run.py scale.jpg -c university.jpg decoder_states/scale/ -n 5

output

Usage

  • In order to used pre-trained decoder weights, set up a directory containing .pth weight files with the form
Relu1_1_decoder_state.pth
Relu2_1_decoder_state.pth
Relu3_1_decoder_state.pth
Relu4_1_decoder_state.pth
Relu5_1_decoder_state.pth

If you want to train your decoders on your own image instead, this directory will be used to store the generated weights.

  • Set layer-specific parameters like the number of epoch and the learning rate for the decoders training phase, as well as iteration parameters in the dictionary observed_layers at the bottom of the file decoders.py.

  • Run the generation process with

python run.py path_to_source_image path_to_decoder_states_directory -o output_path -n n_passes [--train]
  • Run the style transfer process with
python run.py path_to_source_image path_to_decoder_states_directory -c path_to_content_image -s content_strength -o output_path -n n_passes [--train]

Requirements

  • Python 3
  • torch