Skin lesion segmentation without annotated masks must rely on alternative cues to define the target. Existing solutions typically use dermoscopy‑specific heuristics, model priors derived from annotations, or regions selected by users. A general text‑to‑image diffusion transformer offers a textual interface, yet its concept‑word attention is unreliable on dermoscopy images: in the model’s attention output space, the skin‑to‑lesion direction inferred from words poorly aligns with the true separation between real skin and real lesions. We observe that a direction derived from generated healthy‑skin and lesion images aligns much better. Building on this insight, we introduce GENIQ, a training‑free approach that first constructs initial skin and lesion prototypes from generated images, then refines them using the most skin‑like and lesion‑like patches extracted from unlabeled dermoscopy images. Test patches are scored along the refined direction, and a gated threshold suppresses contextual spill‑over around small lesions. Without any inference‑time annotations or extra models, GENIQ attains Dice scores of 0.85‑0.88 on three dermoscopy test sets, outperforming zero‑shot and word‑based attention baselines on all sets, and surpassing a hand‑crafted dermoscopy pipeline in Dice and IoU on two sets.\n\nReview: GENIQ leverages self‑generated images to define a robust feature direction, eliminating the need for manual masks and highlighting a promising new use of text‑image models in medical image segmentation.