jaedukseo/cooc_texture

Co-occurrence Based Texture Synthesis

• Official PyTorch implementation of the paper Co-occurrence Based Texture Synthesis.

Abstract

As image generation techniques mature, there is a growing interest in explainable representations that are easy to understand and intuitive to manipulate. In this work, we turn to co-occurrence statistics, which have long been used for texture analysis, to learn a controllable texture synthesis model. We propose a fully convolutional generative adversarial network, conditioned locally on co-occurrence statistics, to generate arbitrarily large images while having local, interpretable control over the texture appearance. To encourage fidelity to the input condition, we introduce a novel differentiable co-occurrence loss that is integrated seamlessly into our framework in an end-to-end fashion. We demonstrate that our solution offers a stable, intuitive and interpretable latent representation for texture synthesis, which can be used to generate a smooth texture morph between different textures. We further show an interactive texture tool that allows a user to adjust local characteristics of the synthesized texture image using the co-occurrence values directly.

Fidelity and diversity

Interpolation

Naive interpolation in RGB space

Interpolation in the co-occurrence space

Large texture generation with control

Pre-requisites

• Python 3.7
• PyTorch (conda version recommended, pytorch-gpu = 1.3.1)
• torchvision, numpy, scipy, matplotlib, tqdm, pillow, os, scikit-learn

In order to install all the requirements via conda, just run the following command:

conda create -n cooc_texture python=3.7 pytorch-gpu=1.3.1 torchvision numpy scipy matplotlib tqdm pillow=6.1 scikit-learn

and then activate the environment:

conda activate cooc_texture

Training

Run the following command for training:

python train_model.py --texturePath=samples/marbled_0095.jpg --imageSize=128 --kVal=4 

Takes around 3 hours on a NVIDIA 1080Ti GPU, for an image crop size of 128x128 and kVal=4.

• Here, kVal is the number of cluster centers for co-occurrence calculation.
• You can use any image you want, and give the path in the format shown above.

Evaluating

python evaluate_model.py --texturePath=samples/marbled_0095.jpg --modelPath=results/marbled_0095_2020-03-02_23-02-09/ --checkpointNumber=120 --kVal=4 --outputFolder=eval_results/marbled_0095/ --evalFunc=f_d  

Replace the modelPath with the actual path of the directory. The one included is just for representation purposes

• modelPath is the folder containing the checkpoint models.
• checkpointNumber is the number of the epoch of which model we want to use.
• outputFolder is the location where we want to save the evaluated results.
• evalFunc is the selector between:
  • f_d = Fidelity and Diversity - saves an image showing the fidelity with respect to the input crop, and diversity with respect to random noise vector seeds.
  • interp = Interpolation - saves an image of the interpolation between two crops, by interpolating in the co-occurrence space and then generating the result.
  • write_tex = Large Image Generation with Control - saves a high resolution image generated from the input co-occurrence and writes the text '2020' on it.

Check config.py for all options

Acknowledgements

🎉 Thank you for Zalando SE, for providing for the code which served as a base for this project.

👍 The Describable Textures Dataset (DTD) for providing with a highly diverse dataset of images.

License

This project is licensed under the terms of the MIT license (see LICENSE for details).