trained on ID art starting from canto VI for NoobVpred1.0 + anima. there are many like it, but this one is mineactivation tag l1mp1p1<hr /><h3 id="anima-base">Anima Base</h3>new cantos, 8-9-10each picture had 2 prose captions: one without name "a single girl with blonde hair and yellow eyes..." and one with name "Don Quixote from Project Moon..."It overcooked on QOH dress.<h3 id="anima-v3-mid-+-strong">Anima v3 - mid + strong</h3>trained on canto 8-9Mid is more creative, better at facial expression, the pick for style mixing. Also more smudgy and anatomically erroneous.Strong is stable and leans towards whiter lighting. Also rigid.New face crops dataset from later cantos and uses came optimizer<h3 id="canto-8-9-version">Canto 8-9 version</h3>51 epochs on id images from canto 8-9 and face crops from previous cantos.Characters untagged; adding tags fucks up generalization. ABMOG optimizer really helped in making this. Adam and Prodigy refused to pick up the style - mayhaps skill issue.Supposed to be shinier and fancier, but I'm not sure it is. Pruned a bit with lora resize tool to try and counteract repetitiveness. Can't quite say it helped, but didn't seem worse either. -<ul><li>everyone have faust hair</li><li>rigid</li><li>grey, like half the dataset</li></ul>+<ul><li>very awa</li><li>i like it</li></ul>If you want version with less epochs, tell me.

Limbus Company [Style] [NoobVpred + Anima]

anima-limbusnew-l1mp1p1.safetensors

anima base

Files with SHA256: ba1b0c2823f7d50a3750ba3d9f463779fd9e87b5abb6253b59a225f926e66171