Can you please show more colorized examples in bw-color task?

[asetsuna]

It seems that less results was showed for bw2color task than other tasks in the paper and the project page. If result looks promise at most cases, I would like to train the model in large dataset.

[kaihuchen]

@asetsuna If it is of any help, I did a bunch of bw-color experiments which are described here http://www.terraai.org/cgan. One conclusion out of these experiments is that it very much depends on the black-and-white style involved (i.e., the graphical effect that you used to create the black-and-white images), and mapping for one black-and-white style usually does not transfer well to another.

[asetsuna]

@kaihuchen Thank you for your response, your are right , image domain matters. But what I am wondering is more results about the colorization task trained on Imagenet dataset in the paper, this domain is more complicated and I want to know the percentage of fail samples and more decent samples.

[zhoulw13]

@asetsuna @kaihuchen hello, I trained the u-net model to do colorization with lsun-church and cifar-10 dataset, and on the first one-class church dataset, it performs quite well; while in the second multi-classes dataset, it can't give reasonable result for me most of the time. I was wondering that can we make the conclusion that GAN is good at transferring within two assigned style rather than doing more complicated task like general colorization?