Balanced CLIP (1M)
Training CLIP-G took >15KwH of energy, CLIP-L took far less <1KwH
The full negative reinforcement (Cosine Dissimilarity) is available on my huggingface, this was paired with a positive reinforcement (Contrastive Loss) using the full frozen vision model in latent space.
PONY CLIP-L has a further 10 epochs using ASGD for very fine-tuned loss.
Description
FAQ
Comments (11)
Is this CLIP trained for better natural language handling or is it focused on a particular concept? I assume character knowledge has been weakened.
It is trained with natural language, characters might be slightly weekend, but you will have far better SDXL language triggers
@Felldude Just a follow up, should the Illustrious Clip-L be paired with 1 of your previous Clip-G or with the Clip-G posted here?
@nickname45 Feel free to experiment, the CLIP-G posted here is by far the largest training, but the Universal CLIP-G might work better for some cases
"Excuse me, what is the difference between this version and the other version, 'NO MERGE - Universal CLIP (FLUX, SDXL, PONY & illustrious)'?"
About 900k images, both will work with SDXL, PONY, and FLUX but the name triggers match the base model listed.
ClipG for IL in the works? Really like the clipL!
Thank you, but do to the size and training time needed, around a week I doubt I will do an iLLustrious model
thanks for the reply.