The cute Sol Marron with a little more brawn more than brains from the <a rel="ugc" href="https://blackclover.fandom.com/wiki/Sol_Marron">Black Clover</a> series. How to use Use trigger word "slmr" and then prompt away. Her full default outfit is tagged as "crop top, blue capelet, white shorts, white shoes, belt". recommend adding "midriff" and "cleavage" for the best results. "Bare shoulders" also work well with the capelet removed. Her physical appearance tags are pruned, and doesn't need a prompt, but I found adding "green eyes" made them pop out more. Disclaimer: The images used to train this model were a jumbled mix of blurry anime screen shots and two official art images that were not entirely 100% in tune with the show. The initial results were rather poor, but I used generational training to clean up and improve the small dataset. Training a model on its own inputs comes with the risk of inherent flaws due to 'inbreeding', but the result was satisfactory overall. That said, this LoRA is not as perfect as I'd like, but considering the few models for this character, I decided to release it still. A few general issues: The Pony SDXL 'quality' tags may push the output towards a western style, causing the output to deviate from the character design a bit. Most commonly, it can emphasize the "boyish" aspect of Sol's tomboyish looks too much. Adding the "cute" tag can remedy this. Some parts of her outfit didn't have any good tags, most notably the belt mysteriously floating around her hips can be unreliable. Specifying "holster" and/or "pouch" to negatives can help here. The elbow pads can also take different designs depending on what the model decides to draw data from for the specific generation. Naturally, none of this is an issue if you prompt her in different outfits (or none at all).MISC Trained on Pony SDXL at 1024 resolution with a frankenstein'd dataset of 17 images and lots of foul language. All images are untouched txt2img, no inpainting or helper embeddings used for quality transparence, e.g. - things should look even better if you do. The training was done on a small subset of images, as such the output of this LoRA is not an accurate representation neither in quality or style of the source material. Consider buying the official merchandise if you want the good stuff.

Files with SHA256: ee084f9703e877e1255d7e975beed27c35c37cdfd6fb5b621fcbf5791377bd8d