version 2.0 update:I am not satisfied, but it is at least better than version 1.0. The LoRA was created from a dataset containing 71 images with shoes or stockings and 43 barefoot. Heels seem to work better than other kind of shoes and barefoot is also okay. It is not perfect, generating one or two images which are kind of okay from a batch of four. I think it is the most I can make with my hardware and available images and I tried several different approaches. If I had to guess it would need several hundred images more with around 100 epochs to work reliably - which is out of my range.<hr /> This LoRA was a tough nut to crack. I tried to generate images with a person laying on its stomach, the feet up in the air. And Flux really doesn't know what to do with the feet. Bare feet are not that far off, but shoes totally are.I went through several sample sets and approaches, even a rank 32 with 93 images and 300 epochs - at 2.54 it/s on a 4060Ti you can calculate how long that took. And worse: even if that worked, it bled the image out. I cropped the images to only lower bodies and feet prevent the LoRA affecting the faces long before. The last try was only rank 4 and 100 epochs and that seems to work much better.While developing and testing that thing, i have probably seen more deformed feet than any living plastic surgeon. Even so, it works only barely. For shoes you need around 0.9 strength, barefoot a lot less, like 0.5. Unfortunately it is hard to tell in which direction you are off. If you get too many legs or feet or soles in strange places, go lower. Shoes to big either, but deformed shoes is too weak. I trained it on "feet up" "legs up" and "visible soles" because part of the problem seems to be the missing data about soles which are usually not visible.

Feet Up

new bigger dataset, trained with fp8 precision

Files with SHA256: 4b588d273047d590ab9990ac8f843150cd2a32b688c6a8d32e577d0db513ba15