This is the torch implementation for paper "Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis"
This algorithm is for
- un-guided image synthesis (for example, classical texture synthesis)
- guided image synthesis (for example, transfer the style between different images)
- For CUDA backend: choose 'speed' if your have at least 4GB graphic memory, and 'memory' otherwise. There is also an opencl backend (thanks to Dionýz Lazar). See "run_trans.lua" and "run_syn.lua" for our reference tests with Titan X, GT750M 2G and Sapphire Radeon R9 280 3G.
- guided image synthesis
A photo (left) is transfered into a painting (right) using Picasso's self portrait 1907 (middle) as the reference style. Notice important facial features, such as eyes and nose, are faithfully kept as those in the Picasso's painting.
In this example, we first transfer a cartoon into a photo.
We then swap the two inputs and transfer the photo into the cartoon.
It is possible to balance the amount of content and the style in the result: pictures in the second coloumn take more content, and pictures in the third column take more style.
As building Torch with the latest CUDA is a troublesome work, we recommend following the following steps to people who want to reproduce the results: It has been tested on Ubuntu with CUDA 10.
Step One: Install CUDA 10 and CUDNN 7.6.2
If you have a fresh Ubuntu, we recommend Lambda Stack which helps you install the latest drivers, libraries, and frameworks for deep learning. Otherwise, you can install the CUDA toolkit and CUDNN from these links:
Step Two: Install Torch
git clone https://github.com/nagadomi/distro.git ~/torch --recursive cd ~/torch ./install-deps ./clean.sh ./update.sh . ~/torch/install/bin/torch-activate sudo apt-get install libprotobuf-dev protobuf-compiler luarocks install loadcaffe
Step Three: Download Pre-trained VGG Network Pre-trained network:
cd data/models ./download_models.sh
- Most important parameters are '-style_image' for specifying style input image and '-max_size' for resulting image size.
- The content/style images are located in the folders "data/content" and "data/style" respectively. Notice by default the content image is the same as the style image; and the content image is only used for initalization (optional).
- Results are located in the folder "data/result/freesyn/MRF"
- All parameters are explained in "qlua cnnmrf.lua --help".
- Most important parameters are '-style_image' for specifying style input image, '-content_image' for specifying content input image and '-max_size' for resulting image size.
- The content/style images are located in the folders "data/content" and "data/style" respectively.
- Results are located in the folder "data/result/trans/MRF"
- Parameters are defined & explained in "run_trans.lua".
- This work is inspired and closely related to the paper: A Neural Algorithm of Artistic Style by Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. The key difference between their method and our method is the different "style" constraints: While Gatys et al used a global constraint for non-photorealistic synthesis, we use a local constraint which works for both non-photorealistic and photorealistic synthesis. See our paper for more details.
- Our implementation is based on Justin Johnson's implementation of Neural Style.