Trained for PonyXL V6 on khoya-ss XL colab notebook.<hr />Feel free to order your lora through this article: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/6872/commisions-available-18-only">https://civarchive.com/articles/6872/commisions-available-18-only</a><hr />I noticed that Pony XL V6 is already trained on Dross's art style, but it's not quite polished. So I made a dataset large enough for me to make the best LoRA for this style at the moment on Civit. Trigger word is (dross style), and it is the same that is used to trigger the style natively.You still need to use the recommended VAE.For better results, please use the prompting format I used in my showcase, which is tags. Using natural language won't be as flexible as using tags.- What makes this LoRA distinct from the other ones for the same style?<ol><li>It was trained with network dim 34 and alpha 1, which makes it 200+ mb in size; thus retaining more information.</li><li>The dataset used has 304 handpicked, HQ images without artifacts or logos. PonyXL is already trained on samples of this specific artist, and those samples are full of logos and appear in every native generation.</li><li>The overall training was done in 1872 steps and ended with 0.648 loss value which is fairly decent and captures the style, yet being versatile.</li></ol>Note1: Use the LoRA with CFG scale of 3 or 4 and pick DPM++ 3M KERRAS for best results; as the brush strokes' style is grainy and forced the model to consider it realism of some sorts (I couldn't run from this as I used HQ images in training).Note2: You can run the result through a face detailer to fix some eyes in specific face positions like (ahegao).And yes, I plan on making another LoRA for this style, but comic-wise.Feel free to generate and post with this LoRA, and I highly recommend using it locally with something like ComfyUI for a total control over everything.

Files with SHA256: dfbb0d4506b246ba2f408471d19a96e197e65e102e021687f9ffac4bf5b03cae