What is this?This is an experimental LoRa intended for enhancing Chroma 1.0-HD's ability to generate illustrations of furry/anthropomorphic characters using E621 tags and natural language.If you are interested in doing realistic furries then this is not the LoRa for you unfortunately, as it actively worsens Chroma's ability to produce realistic images.<hr />How do I use this?Start with 0.8 strength and use Euler A sampler and Beta scheduler if you want easy good looking results. For Euler A it performs best with a CFG between 4.0 to 7.0 with 5.5 being a good middle ground.For best results you should prompt using primarily simple natural language, with supplemental e621 tags. The order doesn't really matter, but a longer more descriptive prompt tends to produce better quality images.I also added "high quality" and "masterpiece" to captions of images with high and very high aesthetic respectively, as well as "low quality" and "worst quality" to images with low and very low aesthetic, so you can control the quality of output images to a certain degree.ExampleExample positive prompt:<pre><code>(high quality, masterpiece:1.1), absurdres, detailed drawing of a cute raccoon girl with pink and white fur waving at the viewer cutely, oversized t-shirt and earring through her left ear, she has a long tail, spoken heart, framed background, sparkles, owo, stars in background</code></pre>Example negative prompt:<pre><code>(low quality, worst quality:1.1), jpeg artifacts, blurry image, grainy image, extra tail, extra limb, extra finger, detached, disfigured, wrong location, wrong color, wrong size</code></pre><hr />All of my example images are generated at an average resolution of 512 using the fp8_scaled quantized version of Chroma HD-1.0 and Chroma Cache, and then I do 14 more steps in hi-res fix. Your results may vary!

Furry Art Enhancer for Chroma

V3.0 is trained on a revised data set of 186 images with redone captions consisting of natural language &amp; tags.This version also leans into aesthetic trigger words - use "high quality, masterpiece" in positive and "low quality, worst quality" in negative to increase the aesthetics and accuracy of the generated image.I trained with new settings changes based on experimentation and a bit of research:<ul><li>Batch size: 4 -&gt; 2</li><li>Resolution: 768 -&gt; 512</li><li>Learning rate (prodigy): 1 -&gt; 0.5</li></ul>I also now scale all training images to an average resolution of 512 using a ChaiNNer workflow that preserves outlines and hard edges

Files with SHA256: 848b8aea7cfd9a710060e5f10863741906cfd20ef30aba02c17327a692d50c4e