Outfit swaps with and without the LoRA
Loading Images
Three outfit swaps made with the LoRA, then one swap without and with it. Yellow marks every pixel that differs from the original picture.
Without the LoRA, Qwen Image 2.1 redraws the face, the hair and the background along with the clothes, and the picture comes back a few pixels off. With it, only the clothes change.
On 16 swaps the LoRA never trained on, the result sat 0.1 px from the original picture with it and 3.6 px without. Face pixels that changed: 1.7 % against 14.1 %. Background: 0.9 % against 10.8 %.
How it was made, with all the numbers: research notes on Hugging Face.