Paper #77: Repairing Graphic Design Layouts After Content, Asset, and Canvas Updates: A Source-Conditioned Benchmark

Most benchmarks for graphic design evaluate layout generation from given content. In practice, however, a graphic design layout often already exists and should be updated after some desired modification to text, assets, or canvas format. In this setting, the modification breaks only part of the layout, so the affected regions should be repaired while the rest is preserved. Standard one-shot benchmarks fail to evaluate this requirement: prior graphic design methods often redraw the whole canvas, naive train/test splits may leak source layouts or styles, and a single aggregate score can confound the quality of repaired regions and the preservation of unchanged regions. We define this task as post-instantiation layout adaptation and propose a source-conditioned benchmark. Each instance contains a source layout, a modification request, and an unrepaired post-modification layout. We evaluate repairs with four distinct metrics: fidelity, repair quality, source-referenced preservation, and locality. The benchmark contains 117,318 instances drawn from 23,302 source layouts, with 59 held out as a curated subset of difficult, human-verified examples. On this subset, the strongest source-conditioned baseline outperforms the source-agnostic baseline, with higher scores on all four measured metrics. Human validation confirms that the instances require repair and show substantial agreement when multiple valid repairs are allowed. In addition, our analysis shows that train/test splits made at the modification-request level can leak source layouts and styles. Explicitly holding out source layouts and styles from the training set reduces this contamination. Overall, these results show that source-conditioned evaluation is valuable for accurately measuring performance in the post-instantiation layout adaptation setting.