Instructions below. Post some images you made with this lora! We both get buzz when you post. Also, I want to see whatever madness you create.This lora replicates a stylized western art style and is biased towards skinny, toned women. Eventually I may attempt to decouple the style from the subject matter, but for now it's all-in-one. Body type tags are under-learned in the base model, so body type is tricky to decouple from style.<ul><li>The left-most lora version on the page will always be the one I deem best, not just the latest</li><li>Use @l3aven1ng tag to activate</li><li>For the pure look, avoid score tags (score_9, etc)</li><li>Anima prefers high step counts for non-anime styles. I typically use 37 steps in the online generator assuming no distillation. The official Turbo checkpoint also works well (7 buzz a pop in the online generator!).</li><li>Trained on around 75 images in various styles with high quality anatomy and lean female body types. To deal with the variety of styles, I add artist tags to most images and add synthetic (ai generated) data in the lora's target style with my made up artist tag. So far this technique works.</li><li>The name of the lora is a play on my username and has no real meaning.</li><li>Lora training was done through the civit trainer with cosine scheduler. Lower learning rates seem better.</li></ul>

Leavening Style

The result of accidentally training with the wrong parameters in the civit UI

leavening_style_v7_epoch_10.safetensors

v7.5 blooper

v1 extended

Seems more stable, less shiny, and a little "softer" than v5. Also a little muddy somehow? I think v5 still wins

Files with SHA256: 34125f18629f6620d50a5b512fef86bf77a010e151147901ba9ae258e425810c