ModelsLab/ V3X 1216x1216 - β BF16

This one is a full re-haul with heavy training in depiction.

It has data from everything; from hands, to LAION, from models, to cinematic scenes, battles, wars, nsfw, liminal spaces, large landscapes, large character arrays, and many many more new types of data meant to strengthen and augment SDXL.

I took V2-Full and ran 2 epochs of 900k images, instead of many epochs with less images as a curriculum bypass test.

So far the outcome isn't as good, so I'll likely be returning to curriculum training for the DELTA release; but it's generally a bit more responsive than you'd think.

I'll get the exact step count later.