>>63984965Coming back to this, merge results have failed...for now.
Wall of Text incoming!
In case anyone else wants to have some fun:
As merging vpred and eps models was resulting in awful merges I used
https://github.com/arenasys/stable-diffusion-webui-model-toolkit to pull out the Clip model from easyfluff and import it into MeinaHv5.
This would allow a normal SD model take advantage of a possibly better Clip without affecting the model's aspect. The results were mixed and not as earth shattering as I hoped but might turn out something good in the end if used with another model.
As for merges, I tried multiple ways of merging and train-diffing hll, easyfluff, fluffyrock and polyfur(zd_base as well but that always turned out to create very bad results). Add difference between MeinaClipFixed+Easyfluff-PolyFluff seemed to give out the best results in terms of colors not being washed out but they were really sub par when it came to aesthetics. Tried a lot of block merges as well but they didn't come too close to what was being expected. Again, I should probably try this with another more basic sd model and not an overly mixed one.
As a last avenue I inspected the model's metadata and the json format looked awfully similar to something I have seen before. Scrolling down on it revealed that 2/3rds of it was a comfy workflow. Saved it as a Json, fixed brackets, and I can now see how Easyfluff has been merged and tested. Unfortunately somewhere on version 7 there is a merge between the models "1.fp16.safetensors" and "2.fp16.safetensors" with no clear idea of what they might have been (possibly a mix of EasyFluffv6, PolyFur or Fluffyrock, I assume). And here I am now, trying to backtrack this model merging and spin it in another direction but I'm not very good at merging myself.
I didn't really save any image results as they were awful but I'll add one just for the sake of it. I might continue exploring this soon, but now I want to take some time to either fix my kohya or gen a bit.
