Make sure to use UI that supports v-prediction, otherwise you will get a noisy blob!<h3 id="main-parameters:">Main parameters:</h3><ul><li>Steps: 28</li><li>CFG: 3-5</li><li>Sampler-scheduler: euler/euler a-normal, DDPM-exponential (better noses)</li><li>CFGZeroStar and sometimes Self-Attention Guidance: default for difficult backgrounds</li></ul>Try to BREAK prompts into parts by 75 tokens max, this goes for most models..<h3 id="positive-prompt:">Positive prompt:</h3><pre><code>number of characters (1girl/1boy), character name, artist prompt, scene/environment/camera angle, action, expression, items, quality prompts (masterpiece, best quality)</code></pre><ul><li>some heavily stylized characters might require prompt: nose, realistic.. you can put nose into Adetailer</li></ul><h3 id="negative-prompt:">Negative prompt:</h3><pre><code>lowres, worst quality</code></pre><ul><li>even longer negative should work, but I prefer simplicity.. tags worth noting include sketch, abstract, url</li></ul>Enjoy..

Grim Noire vPred

Tried to split my dataset in two, translucency and style. Trained it separately, then I merged it all together in what seemed to be the best way to keep translucency while not overcooking it. (0.666 slimes, 0.25 style) To be clear even slime lora has the same style as the style one, style has just a lot of more images albeit probably a lot of mid ones.. Gotta clean it up i guess, sometime..

Files with SHA256: 374e5b7b5a17c56d5e03a48fbf70ab142bd26afadb495382ede715711425c262