A bidirectional slider LoRA for Krea 2 that dials foot size up or down on fresh generations. It is a slider, not an image editor — it shifts the model's prior, so you set the amount with the LoRA weight and generate normally.

Same seed, foot-size LoRA weight swept −1.5 → +1.6. Only this slider's weight changes across the five frames; the mild background drift is the residual composition leak.

How it works

The effect is a clean bidirectional axis, centered on the base model:

weight 0 — untouched base model (its default foot size)
positive — larger feet; the amount scales roughly linearly with the weight
negative — smaller feet

There is no trigger word. The effect is bound to the LoRA weight, so it is always on — you dial it, you don't summon it with a prompt token.

Usage (ComfyUI)

Load with LoraLoaderModelOnly on Krea 2 Turbo and set the strength:

Recommended range: center 0; useful positive roughly +0.5 to +1.5, negatives to shrink. +0.5 is a mild, natural enlargement; +1 is clearly bigger; toward +1.5 is the top of the clean range.
Fixed seed + strength sweep is the easiest way to find the amount you want for a given image.

Capabilities

Scales across footwear. Confirmed working on bare feet, socks, and shoes — the slider acts on the underlying foot volume, so footwear scales with it rather than clipping. No need to prompt for bare feet.
Bidirectional. Genuinely shrinks as well as enlarges, unlike a one-sided LoRA driven to negative weight.

Limitations

Trained exclusively on male subjects. All training pairs depict men. Behavior on female or androgynous subjects is uncharacterized — the effect may be weaker, less consistent, or may drag in subtly male-coded foot morphology, since the learned axis was calibrated entirely on male feet. Treat results on women as untested.
Scale is tunable, not physical. A given weight corresponds to this build's baked strength, not a fixed anatomical amount. The honest way to know how big a weight looks is to render it.
Trained on Krea-2-Raw; intended to run on Krea-2-Turbo. It transfers cleanly, but was not tuned on Turbo directly.

Training

Method: difference of two LoRAs — a LoRA trained on enlarged-foot images minus a LoRA trained on the matching originals (big − normal).
Data: ~50 pixel-aligned before/after image pairs (normal vs. photoshop-enlarged feet). Captions describe the original scene, so "big feet" never enters the text and the enlargement binds to the LoRA weight instead of a word.
Format: rank-32 PEFT-style LoRA (diffusion_model.*.lora_A/lora_B), ComfyUI-loadable.

License

Released under CreativeML Open RAIL-M. Commercial use is permitted; the license's use-based restrictions (no harm, harassment, deception, exploitation of minors, or protected-class abuse — see Attachment A) apply and must be passed on to anyone you redistribute to.

iinstrum/krea-2-foot-size-slider

Krea 2 — Foot Size Slider