Most of my regular LoRA are trained with &gt; 100 images (ideally). This is my test for a character that only has around 30 usable Garupa arts.<h2 id="heading-3177">Usage</h2><ul><li>This is a LyCORIS/LoCon. Can be used with with WebUI 1.5.0 and up like any other normal LoRA</li></ul><ul><li>Activation tag: <code>okusawamisaki</code> (no space)</li><li>Other recommended tags: <code>black hair, blue eyes, medium hair, bangs, curtained hair</code></li><li>LoRA ratio: should be around <code>0.7</code></li><li>Recommended block weight: <code>1,0,0,0,1,1,1,1,1,1,1,1,1,1,0,0,0</code></li><li>Recommended checkpoint: MeinaHentai v4</li><li>You should inpaint the eyes for better results overall</li></ul><h2 id="heading-3178">Training</h2><ul><li>80 images from Garupa cards (full dataset available <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/ashley-ng/bandori-cards/blob/main/misaki-tagged-cleaned.zip">here</a>). Data pre-augmented using <a target="_blank" rel="ugc" href="https://github.com/deepghs/waifuc">waifuc</a>'s <code>ThreeStageSplitter</code></li><li>Base model: MeinaHentai v4 x0.7 + AnyLoRA AnimeMix x0.3</li><li>Training was conducted in two steps:<ul><li>Pre-train a Bandori art style LoRA with around 2100 card images (dataset available <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/ashley-ng/bandori-cards/blob/main/bandori-cards-tagged.tar.gz">here</a>) w/ around 3000 steps</li><li>Continue Misaki LoRA training from the previous LoRA w/ around 1200 steps</li><li>Both uses the same settings:<ul><li>AdamW8bit, lr scheduler cosine_with_restart (3 restarts), 15% warmup</li><li>Unet lr 3e-4, TE lr 1e-5</li><li>LoCon dim 32/16, conv dim 12/6</li><li>min_snr_gamma=5</li></ul></li></ul>==&gt; The LoRA still burns, especially the eyes</li></ul>

Files with SHA256: 555b21899fbc508791065c060fda4825cf85c45c710ace87d55a08805b9ea117