Read about Gemma CLIP
Works in SD 1.5, SDXL, FLUX, Hunyuan Video
Do not use in PONY or iLLustrious
In most cases I would recommend distilled CLIP over this model
Trained using custom Gemma Vision model at 1024 batch size (64 with 16 gradient accumulation)
Description
FAQ
Comments (16)
I will post why to not use this in Pony/Illu for you, I got you bro.
Maybe they like gradients lol
how to load with sdxl?
how to load with flux?
Dual CLIP loader, for SDXL you need CLIP-G and CLIP-L - For FLUX T5 and CLIP-L
"In most cases I would recommend distilled CLIP over this model" ... so in what cases would you recommend to use this one then?
It seems to have high zeroshot ability's.
Is it just clip L ?
Yes unless I can figure out if the model is truley improving zero shot, and still not loose the long prompt accuracy
with Flux this CLIP generates vector art for me (when I prompt for some photo ...) - can you try to run it with Flux and post the PNG with meta inside?
I did the logic test in flux with no issue, granted I only ran it a few times
Felldude 1 png file pls ;) .... with nodes inside
OliviaRossi https://civitai.com/images/93289118
Felldude TY!!!
What about its advantages? If compare it with Rouwei-Gemma, which is better? And will you make this for Pony and Illustrious later?
Reading that model the creator appears to have trained a high dim lora on the attentions, or possibly the full unet - so completely different approach.
I am still looking into alternate ways to use the vision model, but would not use the method for PONY or iLLustrious as I did here
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.