From footwear lasts to generative AI: a CAD-driven framework for geometry-conditioned technical sketching
Abstract
Generative AI supports creative design, but footwear sketches must match real-last geometry and transfer to CAD-based pattern-making workflows. Existing footwear datasets mainly cover product images, side-view sketches, retrieval, classification, or virtual try-on, without explicitly relating last geometry to technical design lines. This paper proposes a CAD-driven methodological framework for geometry-conditioned technical footwear design and presents ShoeLastSketch, its proprietary dataset-level operationalisation. Each sample pairs a line-free flattened-last input image with a target image containing CAD-derived sketch lines on the same geometry. The resulting dataset includes multilabel semantic annotations in professional footwear terminology and is generated from master models, valid design combinations, and last-to-last transfer procedures. With five core footwear typologies, 75 lasts, and 37,310 paired samples, ShoeLastSketch provides a geometry-conditioned, semantically controlled, CAD-traceable representation of technical footwear design. Downstream experiments empirically demonstrate the dataset’s utility: its structured representation supports learnable attribute recognition and geometry-conditioned sketch generation while preserving predefined technical constraints.
