Krea 2 Foot Size Slider

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.