Everyone's favorite Jyanshi, <a target="_blank" rel="ugc" href="https://mahjongsoul.fandom.com/wiki/Hannah">Hannah</a>! Trained on NoobAI and Illustrious.<h3 id="version-3:">Version 3:</h3>The goals of v3n were to<ul><li>Move to NoobAI as a base training checkpoint (specifically NoobAI EPS 1.1)</li><li>Add new training data compared to my original 2024 dataset</li><li>Improve tagging for official outfits and adornments</li><li>Reduce style and outfit bleed from training images</li></ul>I don't use NoobAI, I use Illustrious checkpoints. Can I use Hannah v3n?Yes! Not only do the two models share some basic architecture internally, but a lot of the popular checkpoints on Civit are often merges that contain both Illustrious and NoobAI DNA. See the gallery images for the various checkpoints (both Illu and NoobAI, including VPred!) that I've used Hannah v3n with.If you primarily generate on CivitAI, you might notice that your access to checkpoints and LoRAs is limited when you're using Hannah v3n. If that's the case, make sure you toggle the following option on:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/237c9748-47a8-414c-8228-94cc08feaf9a/width=525/237c9748-47a8-414c-8228-94cc08feaf9a.jpeg" />as otherwise CivitAI's generator won't let you mix and match checkpoints and LoRAs with different base architectures (Illu and NoobAI).Primary tags:<pre><code>hannah</code></pre>Recommended brain rot tags:<pre><code>1girl, furry female, midriff, (plump:0.5), bare feet

white hair, white fur, striped hair, long hair, four toes (this tag doesn't do anything it's just good luck)</code></pre>Outfit tags:<pre><code>(official costume, white top, crop top, red trim, cargo shorts, brown shorts, thigh bag, hairclip, earring, mouth veil, anklet)

(tube top, blue trim, layered bikini, shawl, hagoromo, white sarong, tassles, detached sleeves, hair flower, hair ornament)

(black dress, criss-cross halter, center opening, pelvic curtain, black choker, thigh strap, red ribbon, flower accessory)</code></pre><h3 id="version-2:">Version 2:</h3>Primary tags:<pre><code>hannah</code></pre>Recommended brain rot tags:<pre><code>1girl, kemono, chubby, bare feet

white hair, white fur, long hair, four toes (this tag doesn't do anything it's just good luck)

furry, anthro (Hannah responds decently well to "kemono" vs "furry". "kemono" gives a flatter face with more anime-style expressions. "furry" gives more of a snout with more western-style expressions)</code></pre>Outfit tags:<pre><code>(main outfit, white top, brown shorts, jewelry, hairclip, mouth veil, anklet)

(blue outfit, tube top, shawl, cloak, hagoromo, sarong, tassles, detached sleeves)

(dress outfit, black dress, center opening, pelvic curtain, black choker, thigh strap, red ribbon)</code></pre><h3 id="version-1:">Version 1:</h3>Primary tags:<pre><code>hannah</code></pre>Recommend brain rot tags:<pre><code>1girl, furry, anthro, kemono, chubby, bare feet</code></pre>Tags for main outfit (white and red crop top with brown cargo shorts):<pre><code>main outfit, white top, brown shorts</code></pre>Tags for third outfit (black cocktail dress):<pre><code>third outfit, black dress, center opening</code></pre>Tags for second outfit (religious blue top and skirt)<pre><code>(this currently does not work, it's embedded in the model but I haven't figured out how to activate it)</code></pre>

Hannah (Mahjong Soul, NoobAI/Illustrious)

<ul><li>Less cooked than V1, resulting in smaller file-size and more off-model adaptability</li><li>Face looks a little more consistent, especially eyes</li><li>Tail is more consistently on-model</li><li>Blue outfit (bond outfit in the game) is now easier to draw out</li><li>A little chubbier by default, therefore reducing the need for high weight (chubby)</li></ul>

Files with SHA256: ea3c522ac5dd70f881a576fca6293ef69550b889c2bbe5e3933594f4ececcc51