For v0.0.0, I trained the LoRA on 60 images in total:<ul><li>6 official illustrations and fan art</li><li>54 in-game screenshots</li></ul>The result can be described as experimental. The character's overall features have been captured, but the following details remain unclear or inconsistent:<ul><li>Horn roots: Both in-game and illustrated depictions show the horns positioned around and intertwined with the bun hairstyle in an accessory-like manner. As a result, the horns may sometimes be generated as detached hair accessories rather than as actual horns.</li><li>Tattoo placement and design: The training images were selected to show the tattoos accurately on the crotch, waist, and back. However, the captioning was fairly rough, resulting in inconsistent reproduction. The number of training steps, or possibly the optimizer and learning rate, also seem insufficient.</li><li>Thigh-high stocking pattern: This is partly due to insufficient training images. In the game, the sides of the thigh-high stockings feature a vertically arranged pattern consisting of two diamond shapes and an arrow. Only a few of the 60 training images clearly show this pattern. As a result, the side pattern is usually missing or rendered ambiguously.</li><li>Belt: In the game, the character wears a thin double belt. Since this detail was not captioned in v0.0.0, the LoRA tends to generate an ordinary single belt instead.</li></ul>

Character > Shelly (kakuno) 小悪魔シェリー（blue arrow garden）

Initial. (Experimental)dim16/alpha16, AdamW8bit, UNet LR 1e-04, cosine 1500, 1024² resolution, 60 images.

Files with SHA256: bf4a2803d7e7e73ed782b9c11515cbf40fc6ee034462bd0f160fa1d05fa943e4