<h3 id="the-preview-images-are-cherry-picked-to-show-the-model-from-its-best-side.intended-to-be-used-only-with-illustrious-models.">The preview images are cherry-picked to show the model from its best side. Intended to be used only with Illustrious models.</h3>An attempt to move <a target="_blank" rel="ugc" href="https://civarchive.com/models/455275/ktf-style-pd">KTF Style</a> to Illustrious. This LoRA's goal is soft style shift, being a blend of comic/cartoon with anime, focused mainly on hair and eyes, but also influencing other facial features and general body shape, especially the head:torso:legs ratio.Much like the PD version, it works best for solo closeups or portraits, and struggles with distant, panoramic or group pictures, with noticeable drops in quality the further the character is from the 'camera'. It's also strongly biased towards feminine appearance, rarely ignoring masculine prompts.Thanks to being an IL model, unlike the PD version, it has better color composition and good lighting quality. However, this style actively goes against some of the strong suits of IL, creating a slightly flatter and chunkier results, with less high frequency details like material texture, small patterns or smooth gradients.It doesn't use any activation keywords, and is meant to be used in 0.8~1.2 strength range. Gets wacky above 1.4, has some weirdness below 0.0, but can go down to -1.0 for interesting results.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1265326?modelVersionId=2406087">MKTF</a> is more stable, flexible and is overall more reliable. <a target="_blank" rel="ugc" href="https://civarchive.com/models/1265326?modelVersionId=1426987">KTF</a> has some minor hiccups with prompt adherence, but offers a more pure style.

KTF Style IL

First full release.A small step up over the test build, merged from dozens of failed training attempts.Compared to KTF:<ul><li>more uniform (albeit somewhat weaker) style</li><li>higher consistency/reliability</li><li>better at small patterns/details (only on closeups)</li><li>(slightly) better at following prompts, with less keyword mixups</li><li>wider variety of hairstyles, facial expressions, poses, clothing and material textures.</li></ul>The goals for future versions are:<ul><li>higher quality =&gt; less prone prone to blurred lines or wonky anatomy</li><li>wider effective range =&gt; style should not fade for keywords outside of training data</li><li>(unlikely) closer style match to the PD version</li></ul>

Files with SHA256: fc6b7c9c9ee5b17a6f373859918468a862fb2bde630cc447272875b0e2378d0a