Is it possible to train image pairs ? (RGB image + Mask image)

[prrw]

Hey,
not sure if it is possible to train the network on paired images ? if not is it possible to have the mask information in the JSON file as label ?

[PDillis]

You could just add the mask as a fourth channel to the image if you want. It's hardcoded that the data should be RGB (num channels = 3), but removing this and making sure everything is saved correctly is a bit tedious but doable. (e.g., when saving a snapshot grid, only save the I did something like this in my repo when you create the dataset for example: https://github.com/PDillis/stylegan3-fun/blob/eab6284c5a556d7540cd708b5ee355679a1795aa/dataset_tool.py#L408