MultiTextureSynthesis

Torch implementation of our CVPR17 paper on multi-texture synthesis.

Prerequisites

• Linux
• NVIDIA GPU + CUDA CuDNN
• Torch
• Pretrained VGG model (download and put it under data/pretrained/)

Task 1: Diverse synthesis

We first realize the diverse synthesis on single-texture. Given one texture example, the generator should be powerful enough to combine elements in various way.

• Training

th single_texture_diverse_synthesis_train.lua -texture YourTextureExample.jpg -image_size 256 -diversity_weight -1.0

• Testing

th single_texture_diverse_synthesis_test.lua 

After obtaining all diverse results, run gif.m (data/test_out/) in Matlab to convert them to an .avi video for view.

To plot the stored training loss (.json file) for any usage,

python plot_loss.py

Task 2: Multi-texture synthesis

• Training

Collect your texture image set (e.g., data/texture60/) before the training.

th multi_texture_synthesis_train.lua

• Testing

We release a 60-texture synthesis model that synthesizes the provided 60-texture set (ind_texture =1,2,...,60) in data/texture60/ folder.

th multi_texture_synthesis_test.lua -ind_texture 24

Task 3: Multi-style transfer

In the synthesis, each bit in the selection unit represents a texture example. In the transferring, we employ a set of selection maps where each map represents one style image when initalized as a noise map (e.g., from the uniform distribution).

Collect your style image set (e.g., data/style1000/) before the training. For large number of style images (e.g., 1000), it is suggested to convert all images (e.g., ,jpg) to a HDF5 file for fast reading.

th convertHDF5.lua -images_path YourImageSetPath -save_to XXX.hdf5 -resize_to 512

• Training

th multi_style_transfer_train.lua -image_size 512

• Testing

We release a 1000-style transfer model that transfers this 1000-style set (ind_texture =1,2,...,1000).

th multi_style_transfer_test.lua 

Citation

@inproceedings{DTS-CVPR-2017,
    author = {Li, Yijun and Fang, Chen and Yang, Jimei and Wang, Zhaowen and Lu, Xin and Yang, Ming-Hsuan},
    title = {Diversified Texture Synthesis with Feed-forward Networks},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
    year = {2017}
}

Acknowledgement

• Codes are heavily borrowed from popular implementations of several great work, including NeuralArt, TextureNet, and FastNeuralArt.