Use this with 
this lora has 105 epochs with 850 different images. its better quality for base than other loras and it can do a lot. there is an emphasis on butts for sure but it can do anything I will start posting images since this isn't gaining traction. get the workflow in my article about combination with base and turbo for quality. i
Description
Turns out this model work good on base as well. But if you start with a two phase worklow that uses base first then refines with turbo its the best of both worlds
in my opinion this works best on zimage turbo...but it was trained on base....I took my original dataset and also created a new one for NSFW sexy times ....the result is a bit frustrating for the XXX stuff but basic women is better than ever. for more complicated stuff sometimes its very good other times it struggles with terminologies. Its an attempt at a general do it all nsfw lora but its only trained on men and women so...perhaps less extensive than that. It can still do very good work. Seems to work best with Euler at about 10-12 steps on zimage turbo..I prefer to run a second pass sometimes using the two pass workflow in my article...it ran for 90000 steps on 800 images. I ran it for 140000 steps total but after 90000 it gets messy and degrades prompt quality. its pretty flexible thanks to being trained on zimage base . This was a learning process for me. But because of the time and money it took to train I won't be updating this for some time. works best for the nitty gritty details at 0.90-1.5 you can run at lower power but it may get genitals wrong or less precisely at lower powers. Feedback is welcome enjoy
FAQ
Comments (11)
Should have secretly trained this on Mr beans face. So that every beautiful woman comes out with Mr beans face as a joke.
That would be a hilarious joke. But I'd have to save it for something less time consuming to train. Best thing to do would be to do a small batch of like 20 photos then do the mister bean trick and tell people its some amazing pony lora hahah
Honestly I wonder everyday if some people sneak in secret trigger words into their LoRAs as jokes that people will likely never find, could you find it?
@iamjustheretodow6140 I'm still new but as I understand it, using a trigger word of any kind could impact a general lora. Even lora's that have trigger words will impact prompts when cranked up. It would be cool if the training process was that exact though
@iamjustheretodow6140 you know before the base model came out I did discover that if you used that de-destilled trainer for zimage turbo and trained a lora at a decent strength if you dropped a rando image into the training data you could re-create it if the lora had trained strong enough, I actually had about 20 or so image in the training data for this that were complete un-related in an effort to ward off over-fitting but when I check for hints of those its pretty weak impact overall. I got lucky because I was tipsy when I started the training job. if I had done that on the de-distilled training adapter this would have come out a style lora. I'm a little saddened by that because it would have been interesting. But apparently training on Base isn't as easy to wobble if you have a few randoms in the data.
Do you like the Z-Image Base version more than the Turbo version? How do you feel about training on Base?
I think training on base is superior for making turbo lora's but generating images in base looks like crap for some reason. I might not be using the right supporting models. I dunno
I"ve been using a workflow that generates zimage images by starting with the base model and finishing with the turbo. The results are better than anything I've seen yet. all my recent posts are done using that format.
@iamgeekusa I definitely have to check that out. Thanks for sharing that.
@Jellai i got the workflow from aitrepreneur
@iamgeekusa hmm I don't wanna give up on zit , does awesome quality just the nsfw out of business .