Learning to Search for Executable Patternmaking from Certified Programs

arXiv cs.GR details SemDraft, a framework for learning executable patternmaking programs from paired sewing patterns without recorded edit steps. The core idea is to reconstruct how one pattern becomes another by searching over editing actions, then validating the result with a deterministic compiler.

That certification step is the practical hook: a recovered program only counts if it recompiles from scratch, its action pairs are classified as ordered or interchangeable, and an independent replay reaches the same pattern. For developers, that kind of checkable transformation is the difference between a plausible output and something you can actually trust in a pipeline.

The paper frames this as a way to move beyond one-shot generation toward editable, stepwise transformations. In game production terms, the same pattern shows up in procedural tools, asset versioning, and any system where artists or tools programmers need to preserve intent rather than just final geometry.

On two frozen draws of 80 order-sensitive cases, the learned ranker certified 63 and 64 cases within 1,000 search executions, versus at most 53 and 52 for non-learned rankers. The result suggests that learned ranking can make bounded search more practical when the search space is large but the output must remain auditable.