<p>Toudou Kirin is a supporting character from the series Gakusen Toshi Asterisk (The Asterisk War). This LoRA is mostly trained on her anime appearance, which aired in 2015-2016.</p><p>I used an experimental method to tag and train this LoRA, aiming for maximum flexibility. It should be very easy to change every aspect about her: hair, clothing, style/colors, body, etc. I have included a technical section at the end for those who are interested.</p><h2 id="usage-iw7843wog">Usage</h2><p>There is no activation word. The main tags are <code>1girl, purple eyes, purple hair, long hair, hair between eyes, two side up, sidelocks, large breasts</code></p><h2 id="outfits-5y2cmjbkf">Outfits</h2><p>Not all outfit tags are necessary. Mix and match to obtain the desired effect.</p><h3 id="school-uniform-wl0q3gbn0">School uniform</h3><p><code>school uniform, blazer, button badge, black bowtie, black skirt, pleated skirt, hip armor, black pantyhose, white boots</code></p><p>Note: The button badge (school emblem) is a lost cause. The SEA sweatshop animators did a bad job, and I'm too lazy to manually fix all &gt;250 instances in the dataset. 😥 Inpainting with <code>close-up, button badge</code> and around 0.7 denoise can fix it slightly.</p><h3 id="accessories-1rit7c5ej">Accessories</h3><p>This is the bag she uses to transport her katana.</p><p><code>weapon bag, over shoulder, strap, holding strap</code></p><p>When viewed from behind, add: <code>tassel, button badge</code></p><h3 id="school-uniform-without-blazer-a4mvix25w">School uniform, without blazer</h3><p><code>black vest, strapless vest, collared shirt, short sleeves, white shirt, black bowtie, black skirt, pleated skirt, hip armor, black pantyhose, white boots</code></p><h3 id="school-uniform-cold-weather-g75x9s6aq">School uniform, cold weather</h3><p><code>aqua poncho, white shirt, long sleeves, black bowtie, black skirt, pleated skirt, hip armor, black pantyhose, white boots</code></p><p>Note: Her white shirt sleeves can sometimes be colored aqua if too much is showing.</p><h3 id="exercisetraining-outfit-ovqbmhypv">Exercise/training outfit</h3><p><code>pink jersey, short sleeves, collarbone, button badge, pencil skirt, hip armor, black leggings, white socks, sneakers</code></p><h3 id="underwear-ownn5p5b6">Underwear</h3><p><code>underwear only, pink bra, bow bra, cleavage, pink panties, bow panties</code></p><p><code>underwear only, blue bra, bow bra, cleavage, blue panties, bow panties</code></p><h3 id="naked-towel-g5bh7xeei">Naked towel</h3><p><code>nude cover, holding towel, hair down, wet hair</code></p><p>Note: Do not use <code>two side up</code> because her hair is down.</p><h3 id="swimsuit-ldbazs7ed">Swimsuit</h3><p><code>white bikini, white sports bra, button badge, cleavage</code></p><h3 id="disguise-1yq7rmhu1">Disguise</h3><p><code>black hat, rabbit hat, blue shirt, dress shirt, blue necktie, blue skirt, frilled skirt, white pantyhose, black boots, knee boots</code></p><p>Note: Do not use <code>two side up</code> when she is wearing her hat because her hat covers it up.</p><h3 id="formal-dress-39r417d10">Formal dress</h3><p><code>black dress, strapless dress, collarbone, cleavage, lace choker, wrist scrunchie, hair scrunchie, strappy heels</code></p><p>Note: There were no detailed shots of her shoes in the anime, but it looks like she wears black strappy heels.</p><h2 id="locations-z9n1fzo3o">Locations</h2><h3 id="arenatraining-room-477h7n9fz">Arena/training room</h3><p><code>arena, audience</code></p><h2 id="swordplay-bjbz1m98y">Swordplay</h2><p>I strongly suggest reading the Danbooru wiki on the relevant tags. The logic is not intuitive.</p><ul><li><p>Katana on hip in sheath, hands not touching it: <code>katana, sheath, sheathed</code></p></li><li><p>Left hand grasping sheath, right hand about to grasp hilt, blade is not visible, popular quickdraw stance: <code>katana, holding sword, sheath, sheathed, ready to draw, iaidou</code></p></li><li><p>Left hand grasping sheath, right hand grasping hilt, blade is visible, partially unsheathed: <code>katana, holding sword, sheath, unsheathing</code></p></li><li><p>Left hand grasping empty sheath, right hand holding katana, fully unsheathed, empty sheath is visible: <code>katana, holding sword, sheath, unsheathed, holding sheath</code></p></li><li><p>Both hands holding katana, fully unsheathed, sheath is not visible: <code>katana, holding sword, two-handed</code></p></li></ul><p>Note: <code>holding sheath</code> is specifically for holding an empty sheath. If the katana is sheathed use <code>holding sword</code> instead, even if her hand is not on the hilt.</p><p>Other useful combat-related tags:</p><ul><li><p><code>fighting stance</code></p></li><li><p><code>battle, clash, sparks</code></p></li><li><p><code>slashing, attack trail, smear frame</code></p></li><li><p><code>speed lines, motion blur</code></p></li><li><p><code>dust cloud, debris</code></p></li></ul><h2 id="nerd-shit-0f8bq2b8m">Nerd shit</h2><p>Here I summarize each step of the process I used to create this LoRA.</p><p>I rewatched the anime with mpv and took many PNG screenshots, pausing and framestepping where necessary. I selected the useful images and with Krita I cropped, rotated, and duplicated/mirrored them. For the manga I chose the highest quality sources I could find then cropped and segmented a few good shots of her on a white background. For the official art I checked Danbooru and the promotional site for the anime.</p><p>Dataset:</p><ul><li><p>955 anime screenshots</p></li><li><p>7 manga segments</p></li><li><p>6 pieces of official art</p></li></ul><p>This is overkill, but I pushed forward just to see what would happen.</p><p>I used TagGUI with SmilingWolf/wd-eva02-large-tagger-v3 to autotag, fixed all the mistakes it made, then added all the tags it missed. This took a long time and made me question the utility of autotaggers. It's better to tag it yourself, assuming you can read the Danbooru wiki. The tagging strategy I used is "keep only the tags you want to learn (delete all unwanted tags)" which gave the most flexibility in my tests.</p><p>I used the following chaiNNer-org pipelines to cleanup the dataset:</p><ul><li><p><strong>Anime</strong>: upscale (2x_AnimeJaNai_HD_V3_Compact) → resize to side (2048, shorter side, upscale only, lanczos) → stretch contrast (keep colors, percentile 0.01) → upscale (2x_AniScale2_ESRGAN_i16_110K) → resize to side (4096, shorter size, downscale only, lanczos) → save (lossless webp)</p></li><li><p><strong>Manga</strong>: upscale (1x_wtp_descreenton_compact) → change color model (rgb to gray) → resize to side (1024, shorter side, upscale only, lanczos) → stretch contrast (keep colors, percentile 0.01) → upscale (4x-DWTP-ds-esrgan-5) → resize to side (4096, shorter side, downscale only, lanczos) → change color model (rgb to gray) → save (lossless webp)</p></li></ul><p>The reason I keep these high resolution intermediates is so I can redraw or inpaint if necessary. Lossless WebP is smaller than PNG, which saves some space. To generate the final training images I use a custom Python script and GraphicsMagick to downscale and crop the dataset into the closest SDXL buckets because I don't trust OneTrainer to do this correctly. An example: <code>gm convert upscale/anime/mpv-shot0833.webp -resize 832x1216^ -gravity center -crop 832x1216+0+0 -define webp:lossless=true ../v1/anime/mpv-shot0833.webp</code></p><p>I used OneTrainer to train the LoRA. Here are my settings:</p><ul><li><p>LoRA: rank 24, alpha 24, dropout probability 0.4</p></li><li><p>Data types: TF32 training, FP32 fallback, FP32 LoRA weight, FP32 base model weight, no overrides, FP16 output</p></li><li><p>Concepts: disable image augmentation, enable tag shuffling</p></li><li><p>Epochs 20, batch size 4, gradient checkpointing, (4760 steps total)</p></li><li><p>Optimizer: Prodigy (beta2=0.99, slice=1, defaults)</p></li><li><p>LR scheduler: constant</p></li><li><p>Text encoder: only train TE1 (CLIP_L), not TE2 (CLIP_G), dropout probability 0.2, stop training at epoch 10</p></li><li><p>Timestep: logit normal distribution, shift 3.15</p></li><li><p>Loss weight function: debiased estimation loss</p></li></ul><p>VRAM usage is 25-30 GB. Training time is 5-7 hours depending on the base checkpoint.</p><p>Rank 16 (possibly lower) seems appropriate for a character with only one outfit, but for a character with several outfits a higher rank was required. Alpha as a hyperparameter is a mistake and should always be equal to rank. Never use BF16 precision because it sacrifices too much quality: it uses 7 bits for the significand compared to 10 bits for TF32/FP16. The relatively high dropout probability is what I think improves quality the most, but it works best with a large dataset and a long training time. To prevent overtraining the text encoder I stop it halfway. Everything else is experimental. Source: it came to me in a dream.</p>

Toudou Kirin (7+ outfits) | Gakusen Toshi Asterisk (NAI,IL,Pony)

Files with SHA256: 825ace6a36cb2031c6135bc7cb3dd531d4207dfd239be998b07de498c612eb45