--updated: V3.0Testing epoch 37. Hoping it's an improvement. One of these days I'll get to training a proper Vpred version somewhere...--updated: V2.0 (example images done on Noobai-XL VPred 1.0, euler_ancestral_cfg_pp, sgm_uniform, cfg 1.4, 28 steps)Lora of character Warning from Fluff-Kevlar. Trained on Noobai-xl Epsilon-Pred 1.1: <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1116447">https://civarchive.com/models/833294?modelVersionId=1116447</a>Should work, possibly at slightly lower quality, on the Noobai-xl Vpred 1.0 too: <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1190596">https://civarchive.com/models/833294?modelVersionId=1190596</a>Reminder that using Eps Loras (the vast majority of them on site) on Vpred models might hurt the Vpred model special qualities. (I'll clarify at the end of the post).<ul><li>Recommended weight: 1.0. It doesn't usually reproduce the blue symbol on thigh correctly though. (improved a bit on V2.0)</li><li>Trigger word: fkwarning<ul><li>Recommended to also add "furry, furry female, 1girl"</li></ul></li><li>Helpful tags (Choose one in each category) :<ul><li>medium hair, short hair, long hair, very long hair</li><li>single hair bun, (any other hairstyle, like, blunt bangs, braided hair, straight hair...)</li></ul></li><li>Style change:<ul><li>flat color</li><li>3d, cell rendering (on V1) // cel rendering (on V2.0)</li><li>combine artists of your choice... just put the artist name, example: "afrobull, neocoill"</li></ul></li><li>Main outfit can be reproduced with:<ul><li>grey dress, sleeveless dress, pelvic curtain, blue vambraces, blue leg wraps, grey thigh strap</li><li>hair ornament, earpiece, neck ring</li></ul></li></ul><hr />Vpred vs Epsilon info:Epsilon models (the vast majority of models on <a target="_blank" rel="ugc" href="http://Civit.ai">Civit.ai</a>) are faulty. They always try to bring the average brightness of the image towards 50%, and will fight your prompt to do so, often adding/removing/changing details and colors of things in the image to make it so. A night scene is usually scattered with lanterns and/or fires, a bright scene might see the main subject clothing's change to dark colors. Also, Epsilon obeys what the seed says there should be in the image, possibly also fighting your prompt with that, which results in seed gatcha.Vpred (Velocity Prediction) models came to fix that! They have better colors and saturation, can make a 100% black/white image, and listen to your prompt way more because of that. That also usually means it won't add all those nice unprompted details that the eps models do. Vpred models are much less affected by the seed, you can usually get the same composition and content on most seeds if you're descriptive enough. When using Vpred models, try lower cfg. They respond strongly.Locally, I recommend Euler Ancestral CFG ++, with SGM_UNIFORM scheduler, and cfg between 1.2 - 2.5 , when using Vpred. 28 steps works fine for me on Noobai. Adding artists to the prompt dramatically impact the result, for better or worse.<hr />If you read all the way here, congrats! Your Iron Will knows no bounds!Happy Gen'ning!

Warning (Fluff-Kevlar) [Noobai-XL]

Improved the dataset: cleaned almost all text and sfx, removed most of the beige backgrounds. Artists and styles are tagged (mostly fluff-kevlar and flat color). Should produce cleaner images than previous versions and support other artist tags better.(Or not. I'm not sure if the training parameters were good... Let's test.)

Files with SHA256: bba202abe28fe0e888fcbffed16f65df70e0c6eb794f124ca48cf8704b3f01a6