Trained on a few illustration someone sent me in the same style (thanks a lot kind stranger). It has not specific author afaik.<h2 id="moss">MOSS</h2>Goodbye Anima, you were fun while you lasted 🥹 Welcome F2-VAE our love story is only just starting.Current DMD/Turbo does not seem to play as well as without with texture heavy LoRA.No trigger trained on this version, though adding something to steer the model in the correct style helps. (e.g <code>Hand-drawn seinen manga illustration, with subtle paper grain and busy lineart. The shading is done with jaggy expressive lines.</code>)MOSS learns extremely well. The dataset is a slim down version of the Anima one composed of literally only 8 images, still manages to generalize the style without collapsing.I might need to think of a way to avoid the facial expression and body shape collapse, but other than that it's a perfect match with the style so I'm really happy.<h2 id="anima">ANIMA</h2>Training texture and fine details is really hard for this model. Unsure if it is caused by Qwen VAE being ass or the Anima architecture/training.It has to be the worse model I've played with in that regard. Still works somewhat if you try hard enough.Trigger is <code>Illustration in a NoirSketch style</code> . Some schizo might help bring out the style more.Example:<code>A light noisy texture is visible all over the anime style illustration. Shading is done through line hatching and crosshatching used for all shading, screentone line pattern in the shadows. The image has washed colors and has dark tones forcing the sharp line shading. film grain. jaggy lines.</code>

Files with SHA256: 97e3aa920f901f4806e637355d523ab208755e04eadd8bd849ad9164b80c5a6c