>>2843789After a fair bit of training on 1.5 and pony I feel like with pony it seems to be a lot more sensitive to captions to bring out the style. Not sure if this was always the case, or just that in 1.5 we were pretty limited in terms of what we could prompt outside 1girls. Or that SDXL inherently sticks to the tokens better, making things harder for lora trainers that don't have very diverse datasets.
Solutions to that would be either adding trigger tokens or training with no captions, neither of which I'm enthusiastic about.
