V3.0 - Improved data set and captioning. First slider LoRA (3rd LoRA overall) — still learning, but pretty happy with this one.I made this mainly because most body weight LoRAs for Flux tend to mess with faces or clash with character LoRAs. This one is designed to avoid that.What I did (dataset):<ul><li>Cropped all faces out of the training images → helps preserve character LoRAs</li><li>All subjects on plain gray backgrounds → doesn’t seem to interfere with scenes</li><li>Tested with multiple LoRAs → plays nicely, especially with character + breast size LoRAs</li></ul><ul><li>All adult women in swimwear</li><li></li></ul>How it behaves:<ul><li>Trained toward the low-noise side, so it’s strong even at low weights</li><li>Higher weight = thinner body</li><li>Sweet spot: -3 to +3</li><li>Outside that range = gets extreme (very thin ↔ very large), around +/-7 it melts down</li><li>Can still generate nudes (likely due to captioning used)</li></ul>Prompting tips:<ul><li>Adds to the effect of the weight</li><li>Use words like: slim, skinny, thick to steer results further</li></ul><hr />If you test it, I’d love to hear how it behaves with your setups 👍All training images were AI generated, no images of real humans were used (living, or deceased).

Files with SHA256: 15849c68bda2cc3af8d7b3eab955c9f470cb74a4779fe1f4e26d59b3218e00e2