wzm2256 / cgCNN

Conditional Generative ConvNets for Exemplar-based Texture Synthesis

This is a tensorflow implement of the cgCNN model proposed in Conditional Generative ConvNets for Exemplar-based Texture Synthesis.

This repository contains implements of three tasks as described below.

Tasks	Description
Texture synthesis	Synthesize new textures that are visually similar to the given exemplar texture.
Texture expansion	Synthesize new textures that are larger (or arbitrarily large) than the given exemplar texture.
Texture inpainting	Fill the corrupted region in the given exemplar texture.

Three types of textures, i.e. dynamic, image and sound textures, are considered for each task, and each application is placed in an individual folder. For example, the code for dynamic texture inpainting is placed in ./dynamic_inpaint, and the code for image texture expansion is placed in ./image_expansion.

Please visit our project page for more results.

Requirements

• python=3.5
• tensorflow=1.8.0
• keras=2.1.6
• librosa (for audio loading and saving)
• ffmpeg (for video loading and saving)

Results

Texture synthesis

Image texture synthesis

Exemplar	Result

Dynamic texture synthesis

Exemplar	Result

Sound texture synthesis

Exemplar	Result
applause_exemplar	applause_synthesis

Texture expansion

Image texture expansion

Exemplar	Result

Dynamic texture expansion

Exemplar	Result

Sound texture expansion

Exemplar	Result
shaking_paper_exemplar	shaking_paper_expansion

Texture inpainting

Image texture inpainting

Exemplar	Result

Dynamic texture inpainting

Exemplar	Result

Sound texture inpainting

Exemplar	Result
bees_exemplar	bees_inpainted

Reference

@article{wang2021conditional,
title={Conditional generative ConvNets for exemplar-based texture synthesis},
author={Zi-Ming Wang and Li, Meng-Han and Xia, Gui-Song},
journal={IEEE Transactions on Image Processing},
volume={30},
pages={2461--2475},
year={2021},
publisher={IEEE}
}

For any question, please contact Ziming Wang ([e-mail]).