Universal CLIP 2
Use Universal CLIP 2 in models like CyberRealistic
Use Universal CLIP 2B in models like RealismStableYogi
DO NOT USE in lucentxlPonyByKlaabu, blendermix or any ponyV6 model that has no SDXL merge in.
Both CLIP-L models should work in FLUX or any setup requiring a clip-vit-large-patch14. CLIP-G should also be a stand in for any CLIP-ViT-bigG-14-laion2B-39B-b160k.
After weeks of work and a few days of training I have decided to release Universal 2, and its clip models without going into Early Access.
Universal 2 has a large corpus of LLM captioned images that have been condensed by T5xxl into 77 token length.
This model understands PONY prompts and characters, SDXL based prompts, and LLM length prompts to an extent.
It does not do text at all.
Description
FAQ
Comments (18)
It looks good in the sample images, but I'm wondering how the result will compare to my 10GB Clip ViT-bigG/14 or my 5GB Flan T5 XXL—not to be argumentative, but because I really don't understand. And could you explain in more general terms which models I should use which version with?
The 10GB CLIP-VIT has the vision model attached, the size is 2.7GB when the vision is pruned, and it should be for most use cases. The CLIP-G and CLIP-L can pair in most SDXL, PONY, SD3.5, models and the CLIP-L in FLUX - I have articles on merging in CLIP, if your unsure.
@Felldude I would be interested in any links to articles on merging in CLIP.
@matari818298279
https://civitai.com/articles/17060/testing-universal-clip-visual-seed-to-seed
https://civitai.com/articles/15600/your-clip-matters-seed-to-seed-comparision-and-workflow
You will need comfy and dual clip loader, trying to use other methods is far more difficult
@Felldude Thank you very much.
Sorry for my ignorance, but which — or all four — of these models can be added to an Illustrious workflow, and how should they be integrated?
Can I load either the L or the G interchangeably — will they work the same way regardless?
I would not use this particular in illustrious, I do have a clip for illustrious that keeps the trained name triggers but aligns it with sdxl
@Felldudehttps://civitai.com/models/2002147?modelVersionId=2272296 — this one. Thanks for your amazing work, bro.
So this is loaded as a normal checkpoint, but only the CLIP output is used, right?
That way I can switch out the regular ones and experiment with yours.
@H_for_Hi The dual clip loader would feed into the clip, the checkpoint or diffusion would feed into model
@Felldude Gotcha! Thanks Fell
Could you please clarify which models these models are specifically adapted for?
Example models are given in the description, however it will work as a drop in for any model that takes base clip-l or G - Illustrious has its own versions of this file which can be found elsewhere
@Felldude Is the "Pony Clip" version of the illustrious version of the file the one from your work?
@Mr_Fei I have not done a universal 2 for illustrious the one is available though - https://civitai.com/models/1784001/no-merge-universal-clip-flux-sdxl-pony-and-illustrious
Sorry for my inexperience, but may I ask if this will only work with Comfy and Forge, or is there a way to use this with reForge as well?
Forge can work, I would test with a seed to seed to make sure - I have had versions of forge not update clip correctly. Comfy UI can be used to create a new checkpoint with the clips
TY for adding some variety to some of the most stagnant parts of the generation!
Thank you