sugi-san / pama

An arbitrary style transfer algorithm

PAMA

​ This is the Pytorch implementation of Progressive Attentional Manifold Alignment.

​ 1/14/2022 Thanks to github user AK391 for making a web demo for PAMA.

​ 1/18/2022 New checkpoints are available. (w/o color loss, 1.5x color loss weight, 1.5x content loss weight)

Requirements

• python 3.6
• pytorch 1.2.0+
• PIL, numpy, matplotlib

Checkpoints

Please download the pre-trained checkpoints at google drive and put them in ./checkpoints.

Here we also provide some other pre-trained results with different loss weights:

Type	Loss	Download
high consistency	w/o color loss	PAMA_without_color.zip
high color	1.5x color loss weight	PAMA_1.5x_color.zip
high content	1.5x content loss weight	PAMA_1.5x_content.zip

The checkpionts will be uploaded recently.

Training

The training set consists of two parts, the content images from COCO2014 and style images from Wikiart.

python main.py train --lr 1e-4 --content_folder ./COCO2014 --style_folder ./Wikiart

Testing

To test the code, you need to specify the path of the content image and the style image.

python main.py eval --content ./content/1.jpg --style ./style/1.jpg

If you want to do a batch operation for all pictures under the folder at one time, please execute the following code.

python main.py eval --run_folder True --content ./content/ --style ./style/

Results Presentation

​ The results prove the quality of PAMA from three dimensions: Regional Consistency, Content Proservation, Style Quality.

Regional Consistency

Content preservation

Style Quality

Other Results