This model was trained on 55 ultra high resolution images from <a target="_blank" rel="ugc" href="http://Hegre.com">Hegre.com</a> (All subjects 18+) I used "Nude woman" for instance prompt and "style" for class. The difference in result from prompting "Nude woman style" + calling the LoRA vs just calling the LoRA is negligible. It doesn't seem to matter as a trigger word.The purpose of the model is to mimic the photographic style of Petter Hegre. Specifically, nude photos of women. However, I've tested it on just inanimate objects like cars, and there is somewhat of a difference in photographic style and composition when the LoRA is turned on and off. Therefore, you can use it for general photography as long as you don't explicitly prompt for "nude woman".<ul><li>Highly recommend using Adetailer by default. Faces are consistently pretty bad, especially the further away the subject.</li></ul><ul><li>RealVisXL produces better results than SDXL, to my eye at least.</li><li>I think it's probably overtrained a particular female subject, of whom there are multiple photos in the dataset. If you prompt "European woman", you will get some of her facial features coming through.</li></ul>v0.9 is epoch 1. v0.99 is epoch 4. v0.99 seems to produce less deformed bodies but worse faces. Try out both. (Sample photos are mixed up w.r.t. epoch and version.)

HegreXL

4th epoch 4,400 steps. Bodies seem to more consistent anatomy. Faces seem to be worse.

Files with SHA256: 63c58502a7e966da1dcbd96b956bf76b62a182e0a8d6a9cde44400e499afa9ea