Discover <a target="_blank" rel="ugc" href="https://aidoll.online">MY ABOUT &amp; LINKS</a> to find more my models and art.This is unique experimental realistic LoRA trained on photos of lovedoll with anime face. The doll name is Akane, she was designed by <a target="_blank" rel="ugc" href="https://twitter.com/Mizuwali">@Mizuwali</a> for <a target="_blank" rel="ugc" href="https://twitter.com/Doll4ever_Ines">@Doll-Forever </a>Irokebijin series.For training I chose mostly SFW Akane photos with blonde hair and blue eyes, but with different skin tones and maybe with a bit different makeups. Also there were some NSFW Akane photos. It should be easy to make her naked. She has beautiful huge breast.Unfortunately, Akane make up with unusual triple eyebrows was not trained perfectly into LoRA (most probably the problem is in dataset, but at this moment it's hard to collect better dataset due to lack of photos). That's why she every-time looks a bit different on renders. If you are curious about how Akane looks in real, here is the link: <a target="_blank" rel="ugc" href="https://www.irokebijinshop.com/index.php/product/irokebijin-silicone-90cm-big-breasts-akane/">https://www.irokebijinshop.com/index.php/product/irokebijin-silicone-90cm-big-breasts-akane/</a>Trigger word: AIDA_LoRA_IrokeAkaneI often use trigger word in constructions like: "close up of (AIDA_LoRA_IrokeAkane:1.13) as beautiful doll" OR "portrait of (AIDA_LoRA_IrokeAkane:1.3) as cute girl".Best Lora values (for exact Akane representation): from 0.8 to 1.05With lower values Akane face most likely will became simplified anime-like or real-like.With higher values you will get artifacts and destroyed images.LoRA trained using kohya_ss on 36 images 768x768 up to 1512 steps with handmade captions and without regularization images. Source model: SD15, learning rate: 0.00001, Precision: fp16, LR Scheduler: constant_with_warmup, LR warmup 30%.

Files with SHA256: 16ba0e7b47f118730ed68fd47bbca6223eee98e9271add62e8e7ad8a3e968d5d