<a target="_blank" rel="ugc" href="https://x.com/d3yk7gh4df">Kunaboto</a> Style Lora."But we've already got 1000 Kunaboto Style Loras!""Good, now we have 1001! I drink whiskey, beer, and coffee. I even eat chocolate."*As with any IllustriousXL or NoobAI XL models, you can add the latent knowledge that the model has of the artist with the corresponding tag. In the case of Kunaboto, that tag is "kunaboto". This will increase the strength of the style, but it may introduce change in detail or colour.<hr /><h2 id="update:">Update:</h2>This model has been, by a large margin, my best performing one. It didn't necessarily need an update, but I've decided to make one. Dataset was refined by removing repetition, clearing censoring bars, getting rid of text bubbles, and adjusting some colors. Also, it was made as a LoCon.Since I had the dataset prepared I ran it through CivitAI's training with their recommended settings for Illustrious.Now that there are more (and better checkpoint mixes) I will also add that version since outputs can be good. Noob still has it ingrained, and I think it's working fine, but for those of you who really want a focus on the style, here it is.The next lines of text were my initial reaction. I leave them, but with a strikethrough so as to be obvious that the statement is no longer true. History should never be erased. Have fun!<hr /><s>The results are not even close to what the Pony-trained one outputs, so I'm not going to upload it</s>. <s>Maybe at a later stage, and maybe with NoobAI as a base, I'd try on my own machine, but for now, the civit's training result gave me pretty much 2-3 very good image results, and all the rest (~20-30) were terrible.</s><hr />

Kunaboto Style [noobV] | [IL] | [PonyXL]

Trimmed dataset for repetition and redundancy. Added some, subjectively (based on my opinion), good new images to the dataset. Main difference - trained on Illustrious-XL v 1.0 - same parameters and training logic.So, if you use "pkkuna" you will get the style from the LoRA. If you use "kunaboto" without the LoRA, you'll get the style from the checkpoint. Using both with the LoRA might bring great results, but occasionally oversaturated images, and force a comic/multiple view images with wonky anatomy.

Files with SHA256: 82d250da79b99c4f0ed42e348133e333d1645a8f412ce70cbae7e0fa0c7296ed