well when i test my loras I always have the same issues: the added stuff looks out of place or the image is super weird but the lora is trained well. So what I learned:

1. if you wanna add stuff (like cum and shit), also prompt: "improve lighting and shadows" - then the cum won't look just pasted over the face but literally as part of the image. cause then the cum also gets "shadows" and is imbedded better.

2. If you don't edit only the person but you also ADD something to the reference image you want to give the AI space in the image. For example my latest Lora with the BBC Deepthroat: If you just have a portrait of the girl that is filling the entire image, give the AI and landscape/horizontal frame to work with. so there is left and right space the add the guy with his schlong. If you keep it in portrait, with will get distorted. You will see, the images with the lora in these showcases are always horizontal but the reference image was always a portrait shot.

3. Currently I am working with portrait stuff. I tried to train an anal-lora with 2000+ images but Flux 2 klein just can't get the girl from a portrait shot into a good doogystyle position or whatever without weird shit happen... currently working on this.

Oh and when you wanna make a lora (your actual question), try to use the automated captioning from civit. It works great, always load up 60 images, captioning them, then the next 60 and so on. Never use the same description on every image. civit will give every image a slightly different caption and with this the lora gets more diverse.

really? I also use drawthigns and it works great. what is the problem?

well I hope all my loras do what they should do haha they all worked well for me on draw things

@aphoticroots well this is exactly how it works. but you need text to image to 100% because you don't want to do inpaint.
Then cfg to 1
Steps to 6
Use uni pc trailing sampler!!!!!
shift to 3.0
Either put the image on the canvas or moodboard that you want to edit.
for NSFW you WILL NEED a NSFW lora like the one I made to generate stuff.

Prompting is highly sensible. After a lora is created I probably take like 1-2h to get the right prompting. Best thing is to use my "triggerwords" as proompting. They will work and then continue from there. Cause else Klein can generate body horror.

Klein does not have NSFW build in. Just prompting doggystyle won't do anything. You will see that my triggerwords are kinda long, because klein need quite the explaination in the prompts. even with lora

Every image I show in my examples has usually just one girl as starting-image.

If you want to generate an image that has the entire body on it the "start-image" should better show the body too.

Thx! Good to be back :)