Is there any further research on clip?

[And233]

I find something interesting in merging clip_g. I use two anime sdxl models, one's dataset is 2024 and the other is 2025. And I only do a weight sum on clip_g with 0.9 between old one and new one. Then the concepts of some 2025's characters are brought to the old one, with no noticeable influence on the style. It seems that clip_l has more knowledge on artist style and clip_g knows more about character.
However, "weight sum" is the only method which can achieve this effect, "subtract" the new one with base model then do "add difference" doesn't seem to be similar. So I'm very curious if you have a deeper understanding of clip merge?

[Enferlain]

Well, you're using a value of 0.9, that's almost full replacement strength. You're essentially taking the entire clip g and overwriting the old one. Just my 2 cents, maybe try 0.5 and see how it goes or compare 0.9 to 1.0

[And233]

yep, I have done many trys, 0.5 maybe too small, some characters need at least 0.7 to see obvious changes. In order to present differences, I use 0.9 for test. Since for the anime model, the amount of knowledge corresponding to a newer dataset is always greater, I don't have to worry about how much of the old character support will be affected by replacing so many. And I'm also convinced that only replacing clipg won't affect most of the artist style representations, because the second model I used has some strange artist style problems and the merge result didn't obviously retain it.
But the main purpose of my conducting this discussion is probably this as well, is there a more reasonable method than weightsum to merge the clip of the two models? It seems that methods like "train difference" are not effective for them.