LoCon trained on Rukisan's artstyle for Pony V7 and Illustrious 0.1Artist: <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/94208188">https://www.pixiv.net/en/users/94208188</a>Check the "About this version" section for any quirks or pitfalls I'm aware of or have experienced.<h3 id="some-q's:">Some Q's:</h3><ul><li>What's your workflow? (Pony)<ul><li>Sampler is used with restarts with segments [3,2,0.06,0.30]</li><li>The scheduler used is Karras</li><li>Initial resolution is selected from the SDXL whitepaper appendix I: <a target="_blank" rel="ugc" href="https://arxiv.org/pdf/2307.01952">https://arxiv.org/pdf/2307.01952</a></li><li>Upscaler is usually 4x_fatal_Anime_500000_G or 4x_NMKD-Siax_200k, with this one in particular upscaling with lanczos also yields acceptable results</li><li>Second sample pass is the same as the first (params are on civit's meta interface) but at ~.35-.45 denoise</li><li>A facedetailer is used</li></ul></li><li>What's your workflow? (Illustrious)<ul><li>Sampler from images is done with nothing special</li><li>No special scheduler is used</li><li>Conditioning is split (A1111 BREAK) into 3 parts, generally BG and camera/Subject Face and attributes/Subject body and clothing</li><li>Initial resolution is selected from the SDXL whitepaper appendix I: <a target="_blank" rel="ugc" href="https://arxiv.org/pdf/2307.01952">https://arxiv.org/pdf/2307.01952</a></li><li>Image is upscaled by 1.5x using Lanczos resampling</li><li>Second sample pass is the same, but at around 15 steps at .4 denoise</li><li>A facedetailer is used after the upscale</li></ul></li><li>What strength should I be using? (pony)<ul><li>Base Pony: From my limited testing ~.8 - 1</li><li>Autismmix: ~.6 - 1</li><li>Other Pony merges: Depends on merge lora compat, but around autismmix levels should work</li></ul></li><li>What strength should I be using? (Illustrious)<ul><li>Lora-compatible merges are fine at .8 - 1</li><li>Otherwise ~.8 or lower</li></ul></li><li>Sample images Model?<ul><li>All Pony images are Autismmix_confetti</li><li>Illustrious images missing the model are <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">cocoIllustriousNoobai_v56</li></ul></li><li>What does your filename structure mean?<ul><li>It's split into 4 parts of [XX]-MAnon-[YYYY]-[ZZZZZZ]<ul><li>[XX] - Artist initials</li><li>MAnon - My moniker</li><li>[YYYY] - SD version and Model version (Ex. XLV4.1 is on SDXL [Not 1.5] and it's version 4.1 of the model, XLILLUS means illustrious specifically)</li><li>[ZZZZZZ] - The epochs, only useful to me</li></ul></li></ul></li></ul><h3 id="explanation-of-issues:">Explanation of Issues:</h3><ul><li>Western cartoon influence<ul><li>The artist tends to draw all characters with big heads and expressive faces, even including ones that should really not. As a result this training tends to lean strongly towards big heads. You can partially remedy this by using "chibi" in the negs at 1+ weight but it's pretty baked in and my attempts to caption it out haven't been successful.</li></ul></li><li>Jagged, crusty linework (seems fixed on Illus version)<ul><li>For the most part the linework is satisfactory, but at points it can be uneven, lumpy, or have inconsistent coloring. I always recommend upscaling, but for this one in particular I recommend upscaling. Tiled upscaling also smooths it out more than usual so it may be something to consider. I think it's the result of downscaling the training images but there's only so much I can do about that with the free time I have.</li></ul></li><li>Issues with angles<ul><li>If you're doing from above/from below/high-angle/low-angle weigh them higher than usual, even then it'll probably fight you</li></ul></li><li>Facial linework has a slight white outline<ul><li>Sometimes the lines around the eyes/mouth I've noticed will have a white tinge around them, which is exacerbated by upscaling beyond 1.5x scale. They're usually easily fixable in paint/ps with a stamp tool but other than than it's something you'll have to live with.</li></ul></li></ul>

Rukisan - LyCORIS LoCon - Pony/Illustrious

Quick retraining of Pony version for IllustriousNot as robustly tested as other trainings - likely to have more issues than listedSimilar issues to Pony version plus:Leans more heavily towards large eyes / smaller nosesConsistency issues - particularly with certain prompts in certain mergesGenerally big headed and bottom-heavy

Files with SHA256: 2a0d740769375443db376de1f67137927c1b167bdef8a85554757c41c4e66f75