<h1 id="heading-17">/!\ New version -&gt; Trained on 20 new images. Same size but better quality /!\</h1>Here is my first trained LoRA for SDXL that enables the generation of black lingerie (which seems to be difficult with the base of sdxl). The 870Mb size of the lora seems to be from sdxl new tensor size (from 768 for sd 1.5 to 2048 for sdxl 1.0) These generated images were not cherry picked at all. You can actually see that a weight of 1 seems to be actually better with generating black lingerie but tend to lower the quality because I trained it on good no head lingerie images but which was not so high in terms of resolution and using a weight of 0.7 is less good on lingerie generation but has more general quality. You can try the lora for any strength between 0.7 and 1. I personally prefer to stays between 0.7 and 0.85.

Black Lingerie - SDXL

More training. Shows a huge difference in terms of quality with the first version.

Files with SHA256: 9a69e4a51ae4a615b5d72570da20464d1c97ab4f298727babe02e4455e836f97