A tensorflow implementation of fast style transfer described in the papers:
- Perceptual Losses for Real-Time Style Transfer and Super-Resolution by Johnson
- Instance Normalization by Ulyanov
I recommend you to check my previous implementation of A Neural Algorithm of Artistic Style (Neural style) in here, since implementation in here is almost similar to it.
All style-images and content-images to produce following sample results are given in style
and content
folders.
Following results with --max_size 1024
are obtained from chicago image, which is commonly used in other implementations to show their performance.
Click on result images to see full size images.
The source image is from https://www.artstation.com/artwork/4zXxW
Results were obtained from default setting except --max_size 1920
.
An image was rendered approximately after 100ms on GTX 980 ti.
Click on result images to see full size images.
- Tensorflow
- Python packages : numpy, scipy, PIL(or Pillow), matplotlib
- Pretrained VGG19 file : imagenet-vgg-verydeep-19.mat
* Please download the file from link above.
* Save the file underpre_trained_model
- MSCOCO train2014 DB : train2014.zip
* Please download the file from link above. (Notice that the file size is over 12GB!!)
* Extract images totrain2014
.
python run_train.py --style <style file> --output <output directory> --trainDB <trainDB directory> --vgg_model <model directory>
Example:
python run_train.py --style style/wave.jpg --output model --trainDB train2014 --vgg_model pre_trained_model
Required :
--style
: Filename of the style image. Default:images/wave.jpg
--output
: File path for trained-model. Train-log is also saved here. Default:models
--trainDB
: Relative or absolute directory path to MSCOCO DB. Default:train2014
--vgg_model
: Relative or absolute directory path to pre trained model. Default:pre_trained_model
Optional :
--content_weight
: Weight of content-loss. Default:7.5e0
--style_weight
: Weight of style-loss. Default:5e2
--tv_weight
: Weight of total-varaince-loss. Default:2e2
--content_layers
: Space-separated VGG-19 layer names used for content loss computation. Default:relu4_2
--style_layers
: Space-separated VGG-19 layer names used for style loss computation. Default:relu1_1 relu2_1 relu3_1 relu4_1 relu5_1
--content_layer_weights
: Space-separated weights of each content layer to the content loss. Default:1.0
--style_layer_weights
: Space-separated weights of each style layer to loss. Default:0.2 0.2 0.2 0.2 0.2
--max_size
: Maximum width or height of the input images. Default:None
--num_epochs
: The number of epochs to run. Default:2
--batch_size
: Batch size. Default:4
--learn_rate
: Learning rate for Adam optimizer. Default:1e-3
--checkpoint_every
: Save-frequency for checkpoint. Default:1000
--test
: Filename of the content image for test during training. Default:None
--max_size
: Maximum width or height of the input image for test. None do not change image size. Default:None
You can download all the 6 trained models from here
python run_test.py --content <content file> --style_model <style-model file> --output <output file>
Example:
python run_test.py --content content/female_knight.jpg --style_model models/wave.ckpt --output result.jpg
Required :
--content
: Filename of the content image. Default:content/female_knight.jpg
--style-model
: Filename of the style model. Default:models/wave.ckpt
--output
: Filename of the output image. Default:result.jpg
Optional :
--max_size
: Maximum width or height of the input images. None do not change image size. Default:None
Train time for 2 epochs with 8 batch size is 6~8 hours. It depends on which style image you use.
The implementation is based on the projects:
[1] Torch implementation by paper author: https://github.com/jcjohnson/fast-neural-style
- The major difference between [1] and implementation in here is to use VGG19 instead of VGG16 in calculation of loss functions. I did not want to give too much modification on my previous implementation on style-transfer.
[2] Tensorflow implementation : https://github.com/lengstrom/fast-style-transfer
- The major difference between [2] and implementation in here is the architecture of image-transform-network. I made it just as in the paper. Please see the supplementary of the paper.
This implementation has been tested with Tensorflow over ver1.0 on Windows 10 and Ubuntu 14.04.