Can I train images with transparency?

[1378dm]

Hello everyone, I want to train some images with transparency, but it doesn't seem to allow me to do so. Does anyone have a way?

[Diego Porres]

In short: there are a lot of changes that need to be done (in particular in dataset_tool.py and training_loop.py), but it's possible. Doing some tests to see that training runs correctly and then generation can be done, so if it's possible, I'll add it to my fork. The starting model snapshot if the dataset has transparency with --cfg=stylegan2:

and with --cfg=stylegan3-t:

So it looks good so far. I will let you know when my tests are complete. Btw, make sure your entire dataset has transparency and not just some, as otherwise, things will break (can't change halfway the dataset type).