Tried training a LORA based on <a target="_blank" rel="ugc" href="https://snickerpuffs.newgrounds.com/art">my own art</a>.<h2 id="current-version-(v2):-0m20ot7cz">Current Version (v2):</h2>Tried tweaking my training method based on <a target="_blank" rel="ugc" href="https://civarchive.com/articles/6503/how-to-easily-create-ponypdxl-lora-better-than-80percent-on-site-just-for-500buzz">randomizer89's training guide</a>, as well as upping my training data from 65 images to 76. Most of these new images are edits of previously-included pieces-- hands and feet I cropped out for previous training attempts, adding variety to the expressions, or even just deleting everything but the background layer.This new model, while a bit looser in some cases (and a bit overfit in others), gets the subtler details of my art style much better than the previous version. I highly encourage you NOT to upscale your results. These look much more style-accurate at lower resolutions, which is how I draw!<h2 id="original-description-(v1):-3wvu385ie">Original Description (v1): </h2>This is actually my third attempt at making this LORA, following <a target="_blank" rel="ugc" href="https://civarchive.com/articles/5758/second-dinners-guide-to-fast-easy-free-high-quality-loras">this guide by SecondDinner</a>. The first one was hand-tagged and, while accurate, was a pain to work with. The second had no tags and was awful all across the board. This third version is... fine? Not perfect, but malleable and close enough, which is really all you can ask from AI.HOW TO USE: The trigger word "sn1ck3erpuff5" should be all you need, but if that's not enough try adding "outlines" and/or "flat colors" to make it more source-accurate. Recommended strength is between 0.8 and 1.2, but if it's giving you trouble you can go as low as 0.6 and still get decently close to the style.Trained on 63 images from the last three years.<ul><li>Most are upper body/'cowboy shot' pieces, focused on solo subjects, with simple backgrounds</li><li>To fill out the training data, I included several unfinished pieces, which were cropped or roughly finished to not be too detrimental to the training process.</li><li>More than ~12% of the training data is furry art, so the model should(?) be able to handle anthro subjects without too much style loss, although I haven't tested this.</li><li>Less than ~7% of the training data is non-feminine (androgynous/male) subjects. Oops.</li></ul>

Files with SHA256: b690cafd4cbe492444aaade1e7aa01e76fb9cfb232fe916677944528cea3a8ff