Background Pony #D9E2
[@Nocturn](/forums/dis/topics/vent-thread?post_id=1419#post_1419)
Nah i train all of my own stuff but for the checkpoints that i dont have the know how or resources to train.
The reason why its such a pain is that to get a clean and consistent style on SDXL architecture is not so easy unless you train it, and on top of that all current SDXL derivatives have heavy anthro bias. So if you train one clean feral pony dataset you pull much better concepts even if they aren't in your dataset.
The moment you try to mix or add stuff to obfuscate the styles. You overlap conflicting things and everything starts to turn into mush. That's why pony diffusion really sucks a certain concepts/details because its been mushed.
Nah i train all of my own stuff but for the checkpoints that i dont have the know how or resources to train.
The reason why its such a pain is that to get a clean and consistent style on SDXL architecture is not so easy unless you train it, and on top of that all current SDXL derivatives have heavy anthro bias. So if you train one clean feral pony dataset you pull much better concepts even if they aren't in your dataset.
The moment you try to mix or add stuff to obfuscate the styles. You overlap conflicting things and everything starts to turn into mush. That's why pony diffusion really sucks a certain concepts/details because its been mushed.