GurilaMash Version 2

block level merge to produce better lora compatibility.

Thanks to Leosam for the idea to merge the SPO lora. It is a welcome addition.

Version 2 is incrementally more flexible than V1. To be clear, GurilaMash is not an attempt at a general purpose model. It's scientifically formulated to make porn. That is what it does. It does it very well. Any other use is incidental.

I have about 20 lora's already trained on this model. I will try to upload at least 1 a day.

People are sleeping on BigAsp

It's basically the Pony of photorealistic explicit content generation

I have a short article with some tips i've picked up.

In case anyone wants to try this at home, here's what I did.

Version 2

Block level merge of SDXXL v3, Anteros v1, then that model was merged with Acorn is Boning v1, then finally BigAsp v1. Then a .3 merge of the SPO lora. BigAsp is 70% of this checkpoint, and this model has purposely inherited it's booru-style prompt method.

Version 1

I did a custom Lycoris extract of SDXXXL then merged it into BigAsp at a .7

Then i did LORA extracts of Acorn Is Boning XL and Anteros XXXL at 128/64, then merged them together at 7/3, then merged that LORA into Bigasp at .7

That's it. I just mashed a few buttons. Acorn has amazing prompt adherence and flexibility, and Anteros was trained on a dataset that suits my aesthetic preferences. SDXXXL is the king of quality output.

My overriding concern, being a complete novice, is I didn't want to fuck with BigAsp's prompting, so doing lora merges seemed the safest bet as i'm not great with block level merges.

If you like what I did, give those guys love if you havn't already. Who am I kidding, if you're here, you've probably already downloaded their models.

I have LORA and LYCORIS extracts of all the major models, and my own dataset which I am training now against BigAsp. I also have other merges done, but this is probably the best. So who knows, maybe I'll upload some more shit in the future.