>>23114
>not necessarily be an SDXL finetune
Everything is SD1.5, now using zerodiffusion as base weights. There have been tries by lodestone which led to TPU sharding, but it wasn't nearly promising enough to warrant switching over.
SDXL itself isn't a well thought-out architecture, it was known for a while the unet wasn't the limiting feature, but the text encoder (and also the VAE a little bit) is the main one, and little improvements were done there. Using a LLM as a text encoder is considered but it's not a priority and it's hard to find a working way.
>HD versions
the HD models are for higher training res, which is mostly for large res upscaling since the base model does fine for usual SD resolutions.
>two different training types "better compatibilty" and "better detail"
It's just a rough way of explaining the effects of vpred zsnr, since using a different prediction type leads to better results but makes existing add-ons for eps-pred (loras, etc..) less effective or broken.
If it's meant as a new standard anime model, then that's not an issue, people will eventually retrain their stuff (which is a good thing since early lora suck ass, holy fuck the abmayo lora is bad)
>if naixl is the one to beat
it was planned and started prototyping way before naiv3, so it wasn't made it with that in mind