Pony CLIP-G 100k

This finetune of CLIP-G and CLIP-L works to align Pony CLIP models back to natural language without breaking the tokens that Astra trained on.

It has been aligned well enough to use in SDXL also.

It can be used with anime models, autism mix, realistic models or even in SDXL.

Why use it? Better name triggers such as celebrity. Better natural language recognition.

How hard was it to train, well it took 68GB to train and required a special script to train Astra's CLIP while still using the CLIP-G Vision blocks.