<h1 id="heading-28037">Mereoleona Vermillion</h1>A.K.A - OneegoleonA.K.A - The Fieriest Redhead<h2 id="heading-28040">How to use.</h2>Plop the Lora on. 0.9-1.0 weight works fine. It should also play nicely with most models and style Loras (as seen in the examples).Use either "Mereoleona", or "Oneegoleon" as the trigger (whichever one is easier for you to spell... Seriously Mereoleona still gets me tripped up from time to time). Capable of both NSFW and SFW (NSFW results near the end of the examples).Please leave a comment or review to show me how it works for you!<h2 id="heading-28041">Outfits</h2>For the Outfits, use these tags, add or remove tags depending on the composition (i.e. a portrait probably won't have the pants visible). She's perfectly capable of wearing just about anything else you throw at her as well. Example images of each outfit is in the images aboveDefault Outfit <pre><code>meroutfit1, red cape, tassel, sarashi, popped collar, white shirt, white pants, blue coat, gold trim, red sash, cleavage</code></pre>Her normal outfit, with the cape and stuffRoyal Knight Uniform<pre><code>meroutfit2, white capelet, red cloak, black turtleneck</code></pre> The Royal Knight uniform, with the really big cloak. This one is the most temperamental of the bunch. Sometimes the cloak generates, sometimes it doesn't. Play around with the weights as needed.Battle Outfit <pre><code>meroutfit3, black turtleneck, torn clothes, sleeves rolled up</code></pre>The Royal Knight Uniform, but just the turtleneck. It's torn up since she's been in battle. Data for this outfit didn't have any look of her legs. Pretty sure she's got a black dress on though, so adding "long black skirt" should work.Movie Outfit<pre><code>meroutfit4, blue jacket, long jacket, open clothes, white shirt, white pants, red sash, gold trim, popped collar, cleavage</code></pre>Kinda like her default outfit, but without the cape. Add "short hair" to complete the movie look<h3 id="heading-28042">Add "fang" to literally every prompt.</h3>Do it. It's not Mereoleona without the snaggletooth and I made extra sure that it generates. It took a lot of work. That being said, I also gave the option to turn it off (since in some circumstances it won't generate properly), hence the tag. Use in conjunction with "fang out" if her mouth is closed.<h2 id="heading-28043">Extras</h2>For some Extras (you might have to weight these higher if they're not showing up).Add "fire" to get some magic fire going on!Add "grimoire" to summon her grimoire so she can really cast some scorching flames. Note: There were only a few images with this in the dataset, so results may vary. Grimoires should pop up, but you might get more than one of them.Add "underlighting" to accentuate the lighting from below for that perfect battle rage look.Speaking of battle rage, don't forget "crazy eyes", "evil smile", "evil grin", and "sanpaku" to really go full-on Calidus Brachium. Or for an added effect, just go full on "blank eyes" for the true battle rage.<h2 id="heading-28044">How it was made</h2>Today, on How it's Made: Mereo-loraThis one was a lot of work. A lot of work that I hope was worth it! When I started this lora, there was only a single model of the character on Civit, and no offense, I didn't think it was very good. Fast forward to now, and there's 2 other ones out ( <a target="_blank" rel="ugc" href="https://civarchive.com/models/104823/mereoleona-vermillion-or-black-clover">this one</a>, and <a target="_blank" rel="ugc" href="https://civarchive.com/models/104867/mereoleona-vermillion-black-clover-character-lora">this one</a>). So it looks like I'm a bit slow out the gate. Hopefully the quality and accuracy make up for it, and shows that mine is the superior version of the best Black Clover waifu. At least mine has the snaggletoothThe reason it took so long, is because of a ridiculous and meticulous training process.<ol><li>First I gathered up 2 separate datasets. One containing (almost) every shot of the character in the anime, and another containing every scrap of fanart I could find of the character.</li><li>I carefully edited each image to be free of any watermarks or signatures, cropped them perfectly, and created another 2 datasets of close-up head crops. One for the fanart, and another for the Official art (like 90% of the shots of the character are headshots in the anime anyway).</li><li>I manually tagged each image for all 4 datasets. Making sure to get as close as possible to the image with the tags. I ended up having to tag about 400 images.</li><li>With these datasets I created 4 loras. A pair trained on images of full body shots (one for the official and fanart), and another pair trained on the head crop datasets. Please let me know if you'd like me to upload either of these!</li><li>For each lora I trained around 18 epochs. So now was the process of narrowing down to each of their best versions. This was largely a process of using the XYZ grid and generating a batch of images for each epoch. And then using an old-fashioned tally sheet to assign "points" to a winner for each set of images. Eventually a winner would make itself known by having a noticeable lead over the others. If there wasn't a clear winner, I'd do another round of testing with the closest set.</li><li>Once winners have been chosen, it's time to merge the loras. Because I wasn't sure which would be the best mix, I had to try a few different types, each at different merge ratios.<ol><li>A merge of the combined fanart body and head and combined official body and head</li><li>A merge of the combo of the fanart and official body, and the combo of the fanart and official head</li></ol></li><li>Another round of testing to find out which merge and which ratio was the best.</li></ol>The ultimate winner turned out to be the official body+head lora (at a ratio of 0.7:0.3), merged with the Fanart body+head (also a 0.7:0.3 ratio). With a final merge ratio of take a guess.... 0.7:0.3 A lot of testing to find out that the perfect merge ratio is 0.7:0.3, but oh well ¯\_(ツ)_/¯Needless to say it all took a while.

Files with SHA256: 1bff4270446327af9709e761ca51c77e76b48fb60e695f3a0eaf2ff82543b75a