PONY Long CLIP (Distilled)
Linear Projection and training to 248 token length
Teacher/Student training of cosine similarity 248 to 77 token length
Note: HiDream CLIP-L is untested, but should be a PONY aligned clip that works in Hi-Dream
Description
FAQ
Comments (10)
This is awesome, love your work on pushing the boundaries of Pony w/ CLIP training! Keep it up <3
Thanks
I will start testing the hidream clip to help out
Awesome, I don't think it will gradient but you never know
Not sure if you mentioned this somewhere else but how does this differ from the UniversalPony or PonyJoy CLIPs? Also which CLIP G would you recommend pairing with this?
Universal-G works well but you could use the FP32 PONY-G also. This model is a distilled model meaning it learn the important tokens in latent space from the 248 model and reduced them to 77 to work with existing diffusion models
can you add a tiny description for us that don't understand what this does and what it's used for?
Not simply, it is part of a model you can load it clip loader