v0.0.0 (Initial Release) This LoRA was trained on <code>anima-basev1</code> using 100 images shared under the "dynamic_pose" tag. The dataset was personally and strictly selected based on "excellent screen composition," "pronounced body twists and extended limbs," and "mid-air, Dutch angles, or non-linear character placements."How it Works &amp; Compatibility: When applied, characters will become "restless," actively performing mid-squats, body twists, and full limb extensions. However, if used with already fine-tuned (FT) models, it may have absolutely no effect due to overlapping training data.Known Issues: There is a known issue where the model forgets previously learned concepts. For example, when this LoRA is active, it forgets how to hold a baseball bat, making it nearly impossible to hold a bat with both hands. (You can see this in the showcase image of the girl with the spiked bat; the left and right hands were never perfectly rendered. Without the LoRA, though reproducibility is low, it would occasionally grip the bat with a perfect set of 5 fingers).

Dynamic pose

Initialdim16/alpha8, unet 1e-05(lion8bit/constant), 100images / 2000 steps

Files with SHA256: 2910496948ec877a05a42654b1656c4f2b5e9f618dac8f6f41e319e09756bbbc