Scan-and-Print: Patch-level Data Summarization and Augmentation for Content-aware Layout Generation in Poster Design

TL;DR

This paper addresses the computational and data-efficiency challenges in content-aware poster layout generation by introducing Scan-and-Print, a compact autoregressive framework guided by a vertex-based layout representation (VLR). The Scan component performs patch-level data summarization to focus perception on informative image regions, while the Print component uses patch and vertex mixup across image-layout pairs to synthesize diverse, plausible samples for robust training. The approach achieves state-of-the-art results on PKU PosterLayout and CGL benchmarks with substantially reduced encoder FLOPs and parameters, and demonstrates effective constrained-generation capabilities for real-world workflows. Overall, Scan-and-Print offers a practical, scalable solution that improves layout quality and generation speed, with strong potential for broader multi-modal design tasks.

Abstract

In AI-empowered poster design, content-aware layout generation is crucial for the on-image arrangement of visual-textual elements, e.g., logo, text, and underlay. To perceive the background images, existing work demanded a high parameter count that far exceeds the size of available training data, which has impeded the model's real-time performance and generalization ability. To address these challenges, we proposed a patch-level data summarization and augmentation approach, vividly named Scan-and-Print. Specifically, the scan procedure selects only the patches suitable for placing element vertices to perform fine-grained perception efficiently. Then, the print procedure mixes up the patches and vertices across two image-layout pairs to synthesize over 100% new samples in each epoch while preserving their plausibility. Besides, to facilitate the vertex-level operations, a vertex-based layout representation is introduced. Extensive experimental results on widely used benchmarks demonstrated that Scan-and-Print can generate visually appealing layouts with state-of-the-art quality while dramatically reducing computational bottleneck by 95.2%.

Figures (7)

• Figure 1: Content-aware layout generation task. (a) Heatmap-based paradigm. (b) Our new efforts: data summarization for efficient image perception and data augmentation for enhanced model generalization, aiming for real-time, robust performance.
• Figure 2: An overview of Scan-and-Print. Preliminarily, (a) represents layout $L$ based on the precise geometric properties, i.e., vertices, and grouping relationship, i.e., underlays, to facilitate the following fine-grained procedures; (b) efficiently 'scans' the input image $I$ to perceive only the few patches suitable for arranging element vertices; (c) 'prints' augmented samples $(\Tilde{I}, \Tilde{L})$ as extra training data by mixing patches and vertices across different pairs within the mini-batch to enhance the generalization ability of the autoregressive model.
• Figure 3: Comparisons of computational cost.
• Figure 4: Comparisons of visualized results on (a)-(e) PKU PosterLayout and (f)-(j) CGL datasets' unannotated test splits.
• Figure A: Comparisons of visualized results of the C $\rightarrow$ S + P constrained generation task on PKU PosterLayout, annotated test split.