<h1 id="iris-xl-eyoaoksjl">IRIS XL</h1>IRIS XL is the result of months of experimenting and fine-tuning to create a model that excels in body anatomy, poses, and stylistic rendering—something I’ve been striving to perfect. This model combines a unique mix of real and synthetic images with a novel training approach to push the boundaries of artistic inference.<h2 id="how-it-was-made:-o96pntxan">How It Was Made:</h2><ul><li>210k Artist Images: I started by training the model on around 210,000 images from booru sources, mostly NSFW, to help it grasp body anatomy and expand its artist/character repertoire beyond just pony-style art. To enhance the tagging quality beyond what’s normally found in booru, I added natural language captions to a huge portion of the dataset. This boosted the model’s ability to interpret and generate more nuanced content.</li><li>50k Pony vs 50k NovelAI: After the base training, I fine-tuned IRIS XL on 50,000 pony images versus 50,000 NovelAI images, both using the exact same prompts. The goal was to teach the model the right way to infer and create art by showing it what NOT to do (pony) versus what it SHOULD do (NovelAI style).</li><li>Dimensional Fine-Tuning: I extracted both fine-tunes with dim512 and applied them to the model, but with a twist: the NovelAI fine-tune was given a weight of 1, and the pony fine-tune a weight of -1. This was key to teaching the model that pony-style inference isn’t what we’re aiming for. The result? A model that’s incredibly good at anatomy and poses, far beyond the base pony limitations.</li></ul><h2 id="why-it's-better:-4u61yqqtf">Why It’s Better:</h2><ul><li>Improved Anatomy and Posing: This technique massively boosted the model's ability to understand human anatomy and how to pose characters naturally, something that was a struggle with previous iterations (if you like my Bunny models, you'll love this).</li><li>Tag-specific LR adjustment: During training, I increased the learning rate for tags that appeared less frequently to help the model stabilize and retain its understanding of rare artists or characters. This way, the model didn’t forget less common tags, ensuring a more robust overall performance.</li></ul><h2 id="the-journey:-cuz03u1cq">The Journey:</h2>I’ve been working on IRIS XL on and off for the last two months using my 4090 GPU, testing, refining, and learning. It’s been a ton of work, but I’m really proud of the result. 💪<h2 id="early-access:-wga215epm">Early Access:</h2>Since this has been such a big project and there’s been a lot of testing involved, I’m opening up early access for 2 weeks. If you don’t have buzz but are really interested in trying it out, send me a DM, and I’ll be happy to let you test it. 😊<h2 id="what's-next-c2edjzwey">What's next?</h2>This is just mainly a test, I'd very much like to replicate this using 10x the dataset - but unfortunately I do not have a H100 cluster just lying around so I will expand slowly and steadily :)

Files with SHA256: 546634e1d2f6e962fe65f2c0bce4af16a61653bec6f075af30b056b2b36bdbcb