Finetune of NoobAI, trained on a RTX 3060 for one full epoch and one partial, using <a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training">this script</a>. Dataset is about 4000 images of varied, anime adjacent art I manually curated and automatically captioned with <a target="_blank" rel="ugc" href="https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3">WD EVA02-Large Tagger v3</a>.Text encoders were frozen and only the UNet was trained with conv1 and conv2 layers skipped. This helps enhance existing features without overfitting. Improvements are mostly backgrounds and artstyles.As a base to train on, a merge of two existing finetunes was used; <a target="_blank" rel="ugc" href="https://civarchive.com/models/956018?modelVersionId=1633911">LS Tiro 3.0</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/1445275/seele-noobai-sdxl?modelVersionId=2281915">Seele 2.1</a> and an unreleased version of <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294/noobai-xl-nai-xl">NoobAI</a>. I did a geometric sum on both of them against the base model with <a target="_blank" rel="ugc" href="https://github.com/ljleb/comfy-mecha">comfy-mecha</a>. Then I added the Cosine A of both results with a ratio of 0.69. The VAE is finetuned; <a target="_blank" rel="ugc" href="https://civarchive.com/models/581286/luna-xlilnai-vae">Luna</a>. Thanks to <a target="_blank" rel="ugc" href="https://github.com/Hysocs/">Hysocs</a> for the memory efficient <a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training/blob/main/optimizer/raven.py">RavenAdamW</a>.

Files with SHA256: 8b73c7168f8d48c13542c94303783e81ca11c5246f25b40afa40b8f30bc41622