The input of Encoder T

[diyiiyiii]

Thanks for your excellent work! I doubt why the preset predictor needs the input of X( content image). Looking forward to your reply.

[minhmanho]

Thanks for the interesting question.

Honestly, the preset predictor does not need the content image X to predict the applied preset as X is always "natural". This is a better way to predict presets. Also, we can use features extracted from a single reference image for stylization. This totally can work.

Nevertheless, I want the encoder T to learn the color transformation from a content image to a reference image so that the extracted features work the best for the given content image (encoder T is aware of the content image). Besides, this design will enable positive pair-wise loss (Given content image X and 2 reference images Z and Z' retouched by the same preset P, the color transformation T(X, Z) should be similar to T(X, Z')).

I hope this will help you. Love to hear if you have any further questions.