purpose of models:
i liked arti as a base and wanted to merge with other stuff so it is more stable and can do many prompts well. i have no experience with merging and did this for fun and they ended up ok in my experience. i can try different mixes if requested.
schizo arti heart netaDPO:
schizo DARE merge of artiwaifu + heart of apple + neta DPO at 0.1 dropout. uploading because someone wanted to test it for themselves. unknown if it actually understands tags from the other two models. i use only artiwaifu tags in images. if you do test it and get good results please upload images it will be very helpful. even if u get bad results please leave a comment so others can see
umbra:
schizo arti heart netaDPO merged with umbra mecha
from others testing it is more stable than the first mix
trif-reweik (removed)
4th tail
arti + heart + netaDPO + 4th tail (pony)
another test for nsfw
Good for simple prompts.
Niche/nsfw can lead to bad anatomy.
reweik
DARE merge 0.1 dropout rate of all three. then with reweik
Update 6/6/2024
Unless someone discovers something then the merges are going to be either continuation of the umbra or 4th tail version.
Update 6/8/2024
The schizophrenia is getting to me. I now think that just the arti+heart+neta is the best. Someone help me...
usage:
Images tend to break if you try to prompt something that your tagged artist has no images of.
I use the Vectorscope extension to lower saturation/contrast because i think some artists/cfg is burned.
You can also try this Euler SMEA DY sampler here. In my experience its more stable (maybe placebo), but generation time is longer than Euler A.
This guide may be useful since base model is Arti. Needs testing.
format of prompt in my images (no negative prompt at all):
by xxx,
general tags (1girl, etc.)
aesthetic, detailed, beautiful color, amazing quality,If you do want to use negative do something simple like:
worst quality, low quality, lowres, artist name, signature, bad anatomy,Credits:
Thanks to all the work of the original models and their creators
Description
artiwaifu + heart of apple + netaDPO + 4th tail DARE dropout 0.1
FAQ
Comments (13)
which one out of four is better and whats the goal of these merges?
i have no idea which is the best. some said umbra is most stable, 4th tail is better for nsfw i think. i have some comments in the description about the merges. im just trying to get NAI level on local because i cant afford getting a subscription right now
i should update description maybe with my initial thought process, but i personally have no idea how DARE merge works. i just did it and it result ended up ok.
@low_channel_1503 ah that explains the name schizo merges lol
bud id like to know which of your merges is better in opinion, and same am also coping with pony since cant afford nai and am planning to not pay to use any stable diffusion model in future, well i want to know which one of your merges is flexible in style shifts and is stable with anatomy nsfw and sfw concepts.
@EBIX from what other people said i would recommend the umbra one or the 4th tail one. if you want nsfw (sex specifically) i think 4th tail will work better. if u have the time and space id appreciate if you test both so i can see which one to keep going. the biggest problem i have testing the models is the negatives. i dont know what negative is good or bad since its so many models.
@EBIX for sfw i think it is usually stable. nsfw is still hit or miss until i find good negative if it exists. i also think the prompt itself is very important. you need like a lot of related prompts together or the model doesnt understand. sometimes with the same prompt i get great result and then the other time its an abomination. try this smea dy sampler here, generation time is longer but i find it more stable. https://github.com/Koishi-Star/Euler-Smea-Dyn-Sampler?tab=readme-ov-file#section1
@low_channel_1503 artiwaifu 2 dropped a few days ago and sanaexl v 1.2 is also coming
@EBIX will try schizo merge with arti v2. not sure what i want to merge it with yet. maybe neta 2.0. any other besides sanae?
Why don't you just merge all of them together? Should round out their strengths and weaknesses, no?
Ill try it out later. Idk why i never thought of that lol. But what merge method would you recommend and what should be model a, etc.?
I would think maybe DARE merging them again might do the trick, but there's this 3x3x3 model on this website which apparently uses Triple Sum so that might work? You could try training diff as well, since the latter two models are both evolution models from the base ArtiHeartNeta so theoretically they should be compatible.
@Jemnite I will see what I can do. Was hesitating doing more with 4th tail because some people said the anatomy was messed up with that one
I have not actually tried the other ones. But ArtiHeartNeta was nice though I haven't tried Arti all that much so I'm not sure how it compares.
@Jemnite Thanks for input
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.