A weight-based slider LoRA for Krea 2 that dials foot size on male subjects. There are no trigger words — load it, set the weight, and generate: 0 = off, positive = bigger, negative = smaller. It works on bare feet, socks, and shoes; footwear scales along with the foot.
Usage
Trained on Krea-2-Raw; intended to run on Turbo.
Built for and tested on men. On female subjects the effect is weak and unstable — that's by design, not a bug to report.
Strength guide (v2)
The relative strength varies by footwear: it's strongest on bare feet, then socks, then shoes. In one measured comparison, matching the same visible size change took roughly bare +0.7 ≈ socks +1.1 ≈ shoes +2.0 — but how smoothly the size responds varies with the other parameters of the image, and the response is not always linear. Treat these reference numbers as a starting point.
Bare feet: +0.75 to +1.0 is the reliable range. Toes are the first thing to degrade at higher weights — reroll if they misbehave.
Socks: clean to about +1.5; around +2.0 the sock volume goes soft.
Shoes: the most headroom — coherent up to +2.0.
Negative weights shrink socks and shoes cleanly. One caveat: on bare feet the shrink direction is much weaker — if you specifically want smaller bare feet, v1 handles that better.
What to expect across a sweep
The sample strips show the same seed swept across weights. Use them to calibrate expectations in two ways: how much the feet change, and how much everything else does. Experience shows that composition drift across a sweep is limited to small incidental changes (décor details, minor framing) rather than large or biased ones.
Versions
v2 (recommended): single-file contrastive build. Bidirectional, much better composition stability across the dial, style-neutral at 0, half the file size.
v1: the original difference-of-LoRAs build, usable range −1.5 to +1.6. Kept mainly for one niche: it shrinks bare feet better than v2 does.
v1 also mirrored on Hugging Face: https://huggingface.co/iinstrum/krea-2-foot-size-slider
Description
The original difference-of-two-LoRAs build (rank 32). Usable range −1.5 to +1.6, default ~+1.0. Shows more composition drift across a weight sweep than v2. Its remaining niche: it shrinks bare feet better than v2 does.
FAQ
Comments (5)
🙏 consider do one for female feet thank you
Sorry, don't have any of the necessary source material.
@werefan any 3d software will do the job.
"DAZstudio3d" is your friend and human centric (and free)
This is probably a trained lora (acting as a slider), not a slider, right? Because I've never seen a slider this large in MB.
It's effectively the delta between two loras, one trained on unaltered images and one on altered ones. Taking the difference isolates the contribution of the edits and allows it to act as a slider.
I think of the term "slider" as a behavioral thing rather than implying a particular implementation.