Layout Generation as Intermediate Action Sequence Prediction

Layout generation plays a crucial role in graphic design intelligence. One important characteristic of the layouts is that they usually follow certain principles. For example, principle repetition emphasizes reuse similar visual elements throughout design. To generate layout, previous works mainly attempt at predicting absolute value bounding box for each element, where such target representation has hidden information higher-order operations like (e.g. copy size previously generated element). In this paper, we introduce novel action schema to encode these better modeling process. Instead values, our approach autoregressively outputs intermediate sequence, which can then be deterministically converted final layout. We achieve state-of-the-art performances on three datasets. Both automatic and human evaluations show generates high-quality diverse layouts. Furthermore, revisit commonly used evaluation metric FID adapted task, observe use different settings train feature extractor obtaining real/generated data distribution, leads inconsistent conclusions. conduct an in-depth analysis settle more robust reliable setting. Code available website.