V2 6.10.2024
Well. I used my earlier LoRA and merged it with couple other ones I've done. After that I still noticed it wasn't working as I wanted it to be, so I merged it once again and set the original model weight lower. As tested it works much better than original one.
Should work with any CP (SD1.5), with different results of course. 😉
V1
This is my first (published) LoRA, it's not perfect at all and I'm going to update it as soon as I get better results with her. It's basicly a test version, just wanted to upload this to see if people can get anyhow good results with her. So far she seems to struggle with hands the most. If you have any hints for training a LoRA model I'm all ears.
I decided to train her with my own merge https://civarchive.com/models/295429 so it gives the best results so far, tested with a few another ones but her outcome is way different with those. (Honestly I hate when people train their models with some "random" not well known model, but yes I did that too). 😂
She is an fictional person, just a wild fantasy.
Annika is a bit goofy and crazy one. She loves to show skin so keep that in mind if you don't want to make NSFW stuff. Younger she was a wild child, now 25. Born with a pair of a wonderful breasts, sparkling personality, blonde but not bimbo at all. ((intelligent woman:1.3))
<<I prefer using weight from 0.5-0.7, of course it depends to a checkpoint used etc.>>
(I prefer using Adetailer with 0.3 weight, her face normally comes quite nice even without. But it makes it bit better, higher values have huge changes)
Description
Works best with CBMX Cartoon Boobs Mix https://civitai.com/models/295429
I prefer using weight 0.5-0.7, depends what model you're using. 😊
FAQ
Comments (6)
If you can, try to train the Lora with a standard checkpoint 1.5. By training it with a specific model (as in your case), you risk having your results only using your model. It should be more flexible. It tends to give too many results NSFW ,If you can, try to insert in the set also not NSFW content because it is quite difficult to dress her. It's nice, I hope in a more flexible update. Good work.
Thank you for your hints, I'll try to update my dataset and also use standard SD 1.5. 😉
@Mr_BlackSheep Try on Civitai Train model realistic 1.5 ;)
@BimbomilloIA Well today I tried using SD1.5 as an checkpoint, was bit overburn. I got plenty safetensors but I haven't had time to look for the best one yet. I'm using Colab and Google Drive, today I run out of time at my free Colab time so I have to test it once again at tomorrow. Also updated my dataset and all the images, hoping it's going to be better next time.
@BimbomilloIA How much time it takes using Civitai Train?
@Mr_BlackSheep 1-3 hour. It depends on the number of images you put. If you have about 50 put 2 repetitions and max 20 epoch. If you have more, make sure that the steps do not exceed 4000 steps otherwise it becomes an overTrain.