> 42 images (maximum I can fit on my GPU with 24Gb memory) Probably I'm completely wrong, but surely scaling down those images, at least for the initial iterat

> 42 images (maximum I can fit on my GPU with 24Gb memory) Probably I'm completely wrong, but surely scaling down those images, at least for the initial iterations, will reduce noise, and give you vastly better batches and faster iteration leading to better training -- then as it converges you would crank up the resolution Note that scaling doesn't have to be cubic, there are algorithms that crop less 'meaningful' pixels; giving you smaller files with the same resolution