TaoD2C-Bench shows visual similarity is not implementation correctness
Researchers from Zhejiang University and Taobao & Tmall introduced TaoD2C-Bench, pairing 2,861 production interfaces with 97,652 expert annotations for component, grouping, alignment and positioning constraints.
Why it matters
Generated UI can resemble a design while selecting the wrong component, breaking container relationships or encoding layouts that fail under changing state. The benchmark separates appearance from implementation requirements.
A closer look
The dataset spans 17 commercial platforms with design images, layer metadata and target component libraries; eight models were evaluated on a reviewed subset of 421 designs. Mean end-to-end F1 was 57.37 for component type, 37.19 for properties, 53.04 for grouping, 23.30 for alignment and 32.22 for positioning. Supplying expert requirements improved every code dimension while average visual scores barely changed, so appearance metrics cannot substitute for implementation checks.
Who this is for
Generative-UI and design-to-code researchers, frontend platforms and design-system teams
Original sources
Revisions & corrections (1)
— Primary arXiv abstract and full HTML paper were read. Submission date, affiliations, dataset scale, evaluated subset, model count, per-dimension results, visual-versus-code finding and licensing limits were verified. This is a distinct industrial UI-code evaluation event not present in the catalog.