/vt/ - Virtual Youtubers

Some setbacks while training styles for SDXL. Seems like adaptive optimizers aren't the easy route for everything as I had some bizarre results. In one case the first epoch captured the style okay but rest of them just dropped the style completely. Adafactor didn't fare any better as it takes very little vram but is extremely slow. Might stick with only AdamW as the results are at least somewhat predictable. Scale Weight Norms for styles is still a bit mystery too as sometimes it helps but in some cases the resulting lora refuses to apply any style if there's other character loras in use.