<h2 id="ani10-experiment">Ani10 🎨 [Experiment]</h2>--[<a target="_blank" rel="ugc" href="https://civarchive.com/articles/20103">Pinboard Updated🥝</a>]--This LoRA is trained on a specific style the dominated the anime scene from 2008 to 2010 think stuff like (Rosario to Vampire, Shakugan no Shana,early High School DXD ...)I'm really fond of that style and i wanted to have it at hand hence this little experiment ..The result was more general then initially intended , as it picked up the concept of anime scene , but hey I still like how it turned out .<hr /><h3 id="usage-tips-(normal)-:">Usage Tips (normal) :</h3><ul><li>I personally use it at the weight of 1 so i recommend that</li><li>the LoRA needs to have its trigger words "<code>nv-ani10, 2010</code>" in the prompt</li><li>reduce the quality tags you're using, just keep "<code>masterpiece</code>" that more then enough as the LoRA is trained to handle it without breaking the style.</li><li>the LoRA only tested on <a target="_blank" rel="ugc" href="https://civarchive.com/models/827184?modelVersionId=1761560">WAI-NSFW-illustrious-SDXL V14 </a>(used in the showcase pics)</li></ul><hr /><h3 id="usage-tips-(advanced)-:">Usage Tips (advanced)✨ :</h3><ul><li>just recently tested [<a target="_blank" rel="ugc" href="https://civarchive.com/models/1201815/nai-xl-vpred10-2d-accelerated?modelVersionId=1925170">NAI-XL vpred1.0 2d accelerated</a>] and it works like a charm with this LoRA</li><li>use with the sampler "Euler Ancestral CFG++" and a "CFG = 1.6"</li><li>activate GFG ReScale and set its value to 0.5</li></ul><hr /><h3 id="about-training-:">About Training :</h3><ul><li>the dataset included I wide range of:<ul><li>expressions</li><li>fantasy / scifi / modern setups</li><li>SFW and NSFW data</li><li>male and female characters</li></ul></li><li>all while trying to achieve balanced distribution ration ratio</li><li>the dataset was cropped manually into sdxl native resolutions for better training fidelity while taking care not to remove essential parts</li><li>the LoRA was trained just enough to get a good effect for its dim (16) , while still having high flexibility</li><li>then some of the LoRA blocks were later tuned based on tests and personal preference</li></ul><hr /><h3 id="usage-tricks-:">Usage Tricks :</h3><ul><li>It seems this LoRA have picked up on concept of tv anime styling , so you can nudge it out of if original dataset if you set the weight to 0.7 and add the year you wanna explore for example 2006 and it might just pull it off ! here are the plot tests for that<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/images/109934477">XYZ Plot tests</a> with the LoRA on/off</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/images/109934470">XYZ Plot tests</a> with the LoRA on/off +"anime screen shot" added</li><li>more tests can be found in this <a target="_blank" rel="ugc" href="https://civarchive.com/posts/24334279">post</a></li></ul></li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/073c45f1-4b70-46ea-a6e2-c39b71e467a1/width=525/073c45f1-4b70-46ea-a6e2-c39b71e467a1.jpeg" />

Files with SHA256: 4d5152b96338d605bd77b25b66591a93e34e6c558b64f2378047805a121270ec