Title: Any2Poster: Cross-Modal and Cross-Domain Poster Generation from Any Source

Abstract:

While visual posters serve as an efficient vehicle for conveying complex information, the field of automated poster creation has struggled with standardized measurement. Current evaluation methods are frequently limited by constraints such as restricting inputs to academic papers, focusing on narrow subject areas, or relying solely on superficial visual resemblance. To address these limitations, we introduce Any2Poster Bench, a comprehensive benchmark designed for any-source poster generation. This benchmark assesses systems based on eight distinct input modalities—including PDFs, URLs, PPTX, DOCX, Markdown, LaTeX, notebooks, and videos—spanning five different content domains.

The evaluation framework combines quiz-based probes to test verbatim factual retention and interpretive comprehension with Vision-Language Model (VLM) assessments. These VLMs judge visual quality, layout aesthetics, readability, content completeness, and logical flow, thereby facilitating a reproducible assessment of both information fidelity and visual communication effectiveness.

To demonstrate and validate this benchmark, we present Any2Poster Agent, an end-to-end reference system. This agent parses heterogeneous source materials, organizes key content, plans poster layouts, renders the final designs, and iteratively improves them through visual feedback. In evaluations on Any2Poster Bench, the Any2Poster Agent secured an average accuracy of 87.25% across all input modalities and 87.28% across content domains.

Furthermore, under PaperQuiz-style evaluations—where previous paper-to-poster agents can be directly compared—Any2Poster Agent significantly outperformed PosterAgent-4o. It raised overall accuracy from the 51.06–51.33% range to 72.58% and increased the density-augmented score from 116–121 to 145.16. Collectively, Any2Poster Bench and Any2Poster Agent offer a reusable evaluation resource and a strong competitive baseline for advancing research in multimodal, domain-general poster generation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC