<h3 id="intro-spiel:-epnf2yvji">Intro Spiel:</h3>Ayy-yo, it's ya boi, reaper (lowercase 'r', no relation)! These two LoRA were a ton of hard work, but also a bunch of fun to make. Feel free to share/merge these, so long as you give, ya boi, credit. They should also be available as characters on the Civitai generator. (&gt;'.')&gt;[&lt;3]<h3 id="disclaimer:-2mu40nx0f">Disclaimer:</h3>This bounty submission was commissioned by, <a target="_blank" rel="ugc" href="https://civarchive.com/user/Void91">Void91</a>. /('.').7<h3 id="support:-w6uughkjk">Support:</h3>My dataset was significantly improved by the use of <a target="_blank" rel="ugc" href="https://civarchive.com/user/AsaTyr">AsaTyr's</a> models and preview images, found here: <a target="_blank" rel="ugc" href="https://civarchive.com/models/154514/fizel-synthesis-twenty-nine-sword-art-online-alicization">Fizel</a> &amp; <a target="_blank" rel="ugc" href="https://civarchive.com/models/154526/linel-synthesis-twenty-eight-sword-art-online-alicization">Linel</a>.There was a surprisingly small amount of official art, fan art, and overall content of these two characters to scrape, and my absolute bacon was saved by these. Please go show their pages some love. (づ๑•ᴗ•๑)づ♡<h3 id="model-usage:-qt6mw72cz">Model Usage:</h3>Both LoRA work well, with Fizel's model being the easier of the two. Trained on basic PonyV6, they should be versatile. With that said, please take a moment to consider my following recommendations: &lt;(-_o)\_.\,,/<ol><li>Both chars were trained on the following structure:{char's name}, source_9, source_anime, {rare token}, {prompt}, {tag1, tag2, ...}Replace name/rare token for Fizel with: fizel &amp; nlwx. For Linel use: linel &amp; pafcPlease take a look at the prompts in the example images for more detailed uses.--Rare tokens were used as an attempt to "capture" any aspects of the characters that the model didn't explicitly store in their name tags, quality tag, and style tag, e.g. their specific dress, armor, etc. Unfortunately, from testing the ever-loving-snot out of these model's, I got mixed results from prompt weighting the rare tokens; they are still needed though, I'm just not sure what info they "stored", exactly.</li><li>Both chars were effectively trained with their armor and dresses intrinsic to them. With that said, the model can mess up certain details like color or accessory placement. I recommend picking from the following set of tags to help:{dress, purple juliet sleeves, red ribbon around neck, wrist cuffs}-- I trained the breastplate to be removable. Either add, breastplate, to prompt for it, or simply leave it out.-- Hilariously, some checkpoints add in dinner plates when you prompt for breastplate. In such cases it helps to add, dinner plate, to the negative prompt.</li><li>Colors may bleed from different prompt elements, e.g. from the dress to the hair/eyes. This issue is less prevalent with Fizel and more so with Linel:-- Fizel may need help using the tags, {blue eyes, short blonde hair}-- Linel tends to respond well to, {gentle grey eyes, twin braids}. Linel's hair was a pain in the butt; I had to regularize her dataset with images of both blondes and brunettes to get that strange kind-of-grey hair she has. On the plus, if your UI allows for prompt mixing/scheduling, then you can take advantage of syntax like, [blonde:brown:0.X] hair, for some mix-ratio 'X'. See example images too.</li><li>The models are a little finicky depending on which checkpoint you use. I recommend that you prompt weight (~0.6) any of the purple-highlighted helper tags in sections 2-3 if they start to give your image trouble. Trained on <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749/pony-diffusion-v6-xl">ponyDiffusionV6XL</a>, use this for best results. Consult the example images for how the LoRA responds to, <a target="_blank" rel="ugc" href="https://civarchive.com/models/288584/autismmix-sdxl">AutismMix</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/345933/mugenmalumix-sdxl">MugenMaluMix</a>, and <a target="_blank" rel="ugc" href="https://civarchive.com/user/Void91">HDA_XL </a>(through Kofi support).<h3 id="legal-and-risk!-63gi2rofp">¡Legal &amp; Risk!</h3>It is prohibited to use this model for any commercial purposes, any illegal scenarios/activities, or in any way which violates Civitai's terms of service. Please generate responsibly! ദ്ദി(ᵔᗜᵔ)</li></ol>

Fizel & Linel (r557)

Initial release: Version 1.0Each char was LoRA trained on a dataset of 32 images using the PonyV6 checkpoint, the OneTrainer GUI/scripts, and the Prodigy optimizer. Both chars were trained for the same number of training steps, in batches of 2, for 50&amp;25 epochs respectively, and with 1:1 balanced regularization using the Waifu Research Department's dataset, available here: <ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/datasets/waifu-research-department/regularization/tree/main">waifu-research-department/regularization at main (</a><a target="_blank" rel="ugc" href="http://huggingface.co">huggingface.co</a><a target="_blank" rel="ugc" href="https://huggingface.co/datasets/waifu-research-department/regularization/tree/main">)</a></li></ul>Linel's set was repeated once to account for the increased blonde/brunette regularization.-- Fizel: (32 x 2/2) x 1 x 50 = 1600 steps-- Linel: (32 x 2/2) x 2 x 25 = 1600 stepsDimension/Alpha: 32/16

Files with SHA256: 683cf4330e1af2bb233bd7832178c6260b6dfb363683a2ecf079b2adf829e85a