Learning to Search for Executable Patternmaking from Certified Programs

TL;DR

SemDraft introduces a framework that learns to reconstruct executable editing sequences between sewing patterns without prior step records. It uses bounded search with three certification checks and trains a learned action ranker that significantly improves success rates in finding valid transformation programs.

Why It Matters

This enables automated, verifiable pattern adaptation in fashion design, reducing reliance on expert manual editing and supporting scalable, reproducible pattern development. The approach bridges the gap between visual or textual pattern generation and the iterative, edit-driven workflow used in real-world patternmaking.

Key Findings

• SemDraft recovers executable editing programs from before-and-after pattern pairs using bounded search and deterministic compilation.
• The framework certifies programs through recompilation, action ordering classification, and independent replay verification.
• A learned action ranker achieves higher success rates—63 and 64 certified cases—compared to non-learned baselines on two test sets.