MeuralPaint

Paint images using famous artistic styles in seconds (or less!)

quickpaint using pre-trained styles (models) or trainstyle a new model on a new style

a TensorFlow implementation for feed forward CNN fast neural style transfer, based on [1-3].

works with different versions of pre-trained TF models, many pre-trained models included (more to come)

Examples

input

styles	outputs	outputs	styles
rain princess			the shipwreck of the minotaur
wave			mosaic
the scream			composition vii
la muse			cubist
udnie			feathers

Usage

QuickPaint

GUI:

Command line:

python quickpaint.py -i [ input (content) ] -o [ output (stylized content) ] -m [ model (style) ] -ma [ mask ] -bl [ blend ]

Example: python quickpaint.py -i inputs/stanford.jpg -o outputs/stanford_cubist.jpg -m pre-trained_models/cubist.model -ma 1 -bl 0.05

required arguments:
-i, --input dir or file to transform (content)
-o, --output destination (dir or file) of transformed input (stylized content)
-m, --model path to load model (.ckpt or .model/.meta) from

optional arguments:
-h, --help show this help message and exit
-d, --device device to perform compute on (default: /gpu:0)
-b, --batch-size batch size for feed-forwarding (default: 4)
-a, --model-arch model architecture if models in form (.model) are used, (default: pre-trained_models/model.meta)
-ma, --mask create binary mask from input (@ 1% of max) and mask output (default: 0 -> no masking)
-bl, --blend multiply the original image with the output using a weighting (factor) – to preserve original colors / contrast (default: 0 -> no blending)

TrainStyle

Command line:

python trainstyle.py -s [ style ] -c [ checkpoint dir ] -o [ output test image ] -i [ checkpoint iterations ] -t [ training set path ] -od [ output test dir ] -cw [ content weight ] -tv [ tv weight ] -sw [ style weight ] - b [ batch size ] -l [ learning rate ] -n [ network path ]

Example: python trainstyle.py -s styles/the_scream.jpg -c checkpoint -o stanford.jpg -od test -cw 1.5e1 -i 1000 -b 20

required arguments:
-c, --checkpoint-dir dir to save checkpoint in
-s, --style desired style image path

optional arguments:
-h, --help show this help message and exit
-t, --train-path path to training images folder
-o, --output output test image at every checkpoint path
-od, --output-dir output test images dir
-e, --epochs # of epochs
-b, --batch-size batch size
-i, --checkpoint-iterations checkpoint frequency
-n, --net-path path to VGG19 network (default data/imagenet-vgg-verydeep-19.mat) -cw, --content-weight content weight (default 7.5)
-sw, --style-weight style weight (default 100.0)
-tw, --tv-weight total variation regularization weight (default 200.0)
-l, --learning-rate learning rate (default 0.001)

Dependencies

• Python 2.7.9
• TensorFlow 0.11.0 >=
• scipy 0.18.1
• numpy 1.11.2

To train:

1) COCO dataset (training data) http://msvocds.blob.core.windows.net/coco2014/train2014.zip

2) VGG19 imagenet weights http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat

Acknowledgements

• Implementation based on fast-style-transfer: https://github.com/lengstrom/fast-style-transfer by lengstrom (engstrom at mit dot edu)

• TF vgg network from: https://github.com/anishathalye/neural-style by anishathalye (http://www.anishathalye.com)

• Pre-trained models acquired from: 1) https://drive.google.com/drive/folders/0B9jhaT37ydSyRk9UX0wwX3BpMzQ 2) https://github.com/junrushao1994/fast-neural-style.tf/tree/master/models

References

[1] Johnson, Justin, Alexandre Alahi, and Li Fei-Fei. “Perceptual losses for real-time style transfer and super-resolution.” arXiv preprint arXiv:1603.08155. https://arxiv.org/abs/1603.08155

check out Justin’s repo for a theano/Lua implementation: https://github.com/jcjohnson/fast-neural-style

[2] Ulyanov, Dmitry, Andrea Vedaldi, and Victor Lempitsky. “Instance Normalization: The Missing Ingredient for Fast Stylization.” arXiv preprint arXiv:1607.08022. https://arxiv.org/abs/1607.08022

[3] Gatys, Leon A., Alexander S. Ecker, and Matthias Bethge. “A neural algorithm of artistic style.” arXiv preprint arXiv:1508.06576. https://arxiv.org/abs/1508.06576