Any2Poster: Any-Source Poster Generation Across Modalities and Domains
Any2Poster enables cross-modal, domain-general poster generation with 87.25% accuracy, integrating unified parsing, adaptive layout, and VLM-guided visual repair.
Key Findings
Methodology
This paper introduces the Any2Poster framework, which combines a unified parsing module, content-aware layout planning, and HTML/CSS-based rendering, leveraging multimodal pretrained models such as Visual-Language Models (VLMs). The system comprises six main stages: heterogeneous source parsing (supporting PDFs, web pages, PPTX, DOCX, Markdown, LaTeX, notebooks, videos), content analysis and importance estimation, layout prediction, visual element selection, HTML/CSS rendering, and VLM-guided local visual repair. The parsing module normalizes diverse formats into a structured representation, enabling consistent downstream processing. The layout planner dynamically assigns panel roles and visual styles based on content importance and type, ensuring adaptive and coherent design. The rendering component constructs editable HTML/CSS, allowing precise control over typography, spacing, and visuals. The VLM-based repair loop detects layout issues such as overflow, misalignment, or scale problems, and applies targeted edits, improving visual quality. Evaluation on the newly proposed Any2Poster Bench, which covers 8 input modalities and 5 content domains, demonstrates the system's robustness and generalization, achieving an average quiz accuracy of 87.25%.
Key Results
- Across eight input modalities, Any2Poster Agent achieves an average quiz accuracy of 87.25%, outperforming GPT-4o (74.00%) and Paper2Poster (58.30%). Notably, it exceeds 93% accuracy on Markdown and LaTeX inputs, indicating strong modality adaptability.
- In cross-domain evaluation, the system maintains high performance with an average accuracy of 87.28% across research, news, education, business, and fiction sources, surpassing baseline models significantly.
- In PaperQuiz-style assessments, the overall accuracy improved from approximately 51% (prior models) to 72.58%, with a density-augmented score reaching 145.16, confirming enhanced information fidelity and visual coherence.
Significance
This work addresses a critical gap in automated visual communication by enabling a single system to handle diverse content sources and domains, a feat previously limited to narrow formats like academic PDFs. It advances multimodal understanding and layout synthesis, making automatic poster generation more practical and scalable. The introduced evaluation framework, combining quiz-based information recovery and visual quality assessment, sets a new standard for measuring multimodal content-to-visual transformation. The approach has profound implications for scientific dissemination, corporate reporting, and educational content creation, potentially transforming how complex information is visually communicated at scale.
Technical Contribution
The main technical contributions include the design of a multimodal unified parser capable of converting heterogeneous formats into a shared structured representation, a content-adaptive layout planning algorithm that dynamically allocates visual and textual components based on importance, and an HTML/CSS-based rendering pipeline that supports precise, editable visual output. The integration of VLM-guided local repair enables fine-grained correction of layout issues, reducing hallucinations and visual artifacts. This combination of structured understanding, adaptive design, and visual refinement constitutes a significant step forward in end-to-end multimodal content synthesis.
Novelty
This research is the first to support a truly multimodal, domain-general poster generation system capable of ingesting eight different source formats and producing high-quality visual summaries. Its core innovation lies in the seamless integration of structured parsing, content-aware layout planning, and code-based rendering, enhanced by VLM-guided visual repair. Unlike prior works limited to scientific PDFs or templates, this system demonstrates broad applicability across diverse real-world sources, setting a new benchmark for multimodal content synthesis.
Limitations
- The system's performance diminishes when source content lacks sufficient structure or contains excessive noise, such as poorly scanned documents or unorganized videos, indicating robustness issues with highly unstructured data.
- Support for dynamic or interactive visual elements remains limited; current rendering focuses on static images and text, which may restrict use in interactive or multimedia-rich contexts.
- Computational costs are high, especially during multi-stage parsing and VLM-guided repair, posing challenges for real-time or large-scale deployment without further optimization.
Future Work
Future directions include enhancing robustness to unstructured and noisy data, extending support for interactive and multimedia content, optimizing the pipeline for faster inference, and incorporating user feedback for personalized customization. Additionally, exploring reinforcement learning for layout optimization and expanding multilingual capabilities will broaden system applicability. These efforts aim to make multimodal, domain-general poster generation more efficient, versatile, and user-friendly.
AI Executive Summary
In an era overwhelmed by information, the ability to distill complex, multi-source content into clear, engaging visual summaries is increasingly vital. Traditional methods for creating posters or visual abstracts rely heavily on templates or manual design, which are time-consuming and limited in scope. Existing automated approaches often focus solely on scientific papers or single formats, leaving a significant gap in versatility and scalability.
This paper introduces Any2Poster, a comprehensive framework designed to generate high-quality posters from a wide array of heterogeneous sources. The system integrates a multi-stage pipeline: first, a unified parser converts diverse formats such as PDFs, web pages, PPTX, DOCX, Markdown, LaTeX, notebooks, and videos into a structured representation. Next, content analysis and importance estimation guide the adaptive layout planning, dynamically assigning roles to different content blocks based on their significance and type. The core of the system is a code-based rendering engine that constructs editable HTML/CSS layouts, enabling precise control over typography, spacing, and visual elements.
A key innovation is the incorporation of Visual-Language Models (VLMs) to guide localized visual repair. After initial rendering, the system crops panels and uses VLMs to detect issues like overflow, misalignment, or scale problems. Targeted HTML/CSS edits are then applied, refining the visual coherence without regenerating the entire poster. This iterative process ensures high-quality, visually appealing outputs.
The authors evaluate their approach on the newly proposed Any2Poster Bench, which covers eight input modalities and five content domains. Results show an average quiz accuracy of 87.25%, outperforming existing models such as GPT-4o and Paper2Poster. The system also excels in cross-domain generalization, maintaining high performance across research, news, education, business, and fiction sources.
Furthermore, in PaperQuiz-style assessments, the system's overall accuracy increased from around 50% to 72.58%, with a density score reaching 145.16. These metrics demonstrate the system's ability to preserve factual information and produce visually coherent posters. The work signifies a substantial step toward fully automated, versatile content visualization, with promising applications in scientific dissemination, corporate communication, and education.
Despite these advances, challenges remain. The system struggles with highly unstructured or noisy sources, and the computational cost is significant. Future work will focus on improving robustness, speed, and support for interactive multimedia content, aiming to make multimodal poster generation more accessible and adaptable for real-world deployment.
Deep Analysis
Background
The evolution of multimodal AI has significantly advanced content understanding, with models like CLIP, ALIGN, and Flamingo enabling cross-modal retrieval and understanding. Prior work in automatic poster generation, such as PosterBot, Paper2Poster, and PosterGen, primarily targeted scientific papers, relying on template-based layout and rule-driven visual design. These systems excel in structured, domain-specific contexts but lack flexibility for diverse sources. Recent developments in document AI, including PubLayNet and DocLayNet, support layout analysis and structural understanding across formats. However, integrating these advances into a unified system capable of handling multiple formats and content domains remains an open challenge. The need for a versatile, end-to-end solution that can process heterogeneous sources and produce high-quality visual summaries has driven recent research, but comprehensive solutions are still lacking.
Core Problem
The core problem addressed is how to develop a universal, scalable system capable of transforming a wide variety of source content into visually appealing, information-rich posters. Key bottlenecks include: • Heterogeneous content parsing: different formats (PDF, web, PPTX, etc.) require format-specific parsers; • Structural understanding: extracting salient information without losing context; • Layout planning: dynamically designing coherent, importance-aware arrangements; • Visual rendering: generating high-fidelity, editable visual layouts; • Visual refinement: correcting layout issues and visual inconsistencies. Existing methods are limited by format dependency, narrow domain focus, or lack of adaptive design, making it difficult to generalize across real-world scenarios.
Innovation
This work introduces several key innovations: 1) Multi-format unified parser: supports diverse input formats, normalizing them into a shared structured representation, enabling downstream processing; 2) Content-aware layout planner: dynamically assigns visual roles and arranges content based on importance and type, ensuring adaptive and contextually relevant design; 3) Code-based rendering pipeline: uses HTML/CSS for precise, editable visual output, facilitating fine-grained control and local adjustments; 4) VLM-guided visual repair: employs visual-language models to detect and correct layout issues such as overflow, misalignment, and scale imbalance, iteratively refining the poster. These innovations collectively enable a robust, flexible system capable of handling real-world, multi-source content for high-quality poster generation.
Methodology
- �� Heterogeneous source parsing: utilize dedicated parsers for each format (PDF, webpage, PPTX, DOCX, Markdown, LaTeX, notebooks, videos), converting them into a normalized ParsedDocument schema comprising sections, figures, tables, and raw text; • Structural normalization: split long sections into manageable chunks (~800 tokens), maintaining paragraph and equation integrity, and storing local context; • Content importance estimation: analyze parsed content to predict title, core message, and section significance; • Layout planning: assign panel roles, determine visual styles (prose, bullets, figures), and decide on visual preservation or synthesis based on content quality and relevance; • Visual element selection: retain source figures if suitable, or generate new visuals using prompts derived from content concepts; • HTML/CSS rendering: convert planned layout into editable HTML blocks, embed figures as data URIs, and assemble the complete poster; • Visual refinement: crop panels, detect issues via VLM, and apply targeted HTML/CSS edits for layout correction; • Output generation: export final posters in PDF and PNG formats, supporting user interaction and further editing.
Experiments
The experimental setup involves evaluating the system on the newly proposed Any2Poster Bench, which includes approximately 160 sources spanning 8 modalities and 5 domains. For each source, the system generates 20 factual and 20 interpretive multiple-choice questions, answered by VLM-based readers to assess information fidelity. The evaluation compares the proposed system against baselines such as GPT-4o, Gemini 2.5 Flash, GPT-5, and prior paper-to-poster models like Paper2Poster. Metrics include quiz accuracy, visual quality scores from VLM judges, and density-augmented scores. Ablation studies analyze the impact of each component, and cross-modality and cross-domain evaluations verify robustness. The experiments demonstrate the system’s ability to generalize across formats and content types, with detailed statistical analysis confirming significant performance improvements.
Results
The system achieves an average quiz accuracy of 87.25% across all modalities, with over 93% on Markdown and LaTeX inputs, indicating excellent modality adaptability. Cross-domain accuracy remains high at 87.28%, outperforming baseline models by a substantial margin. In PaperQuiz-style evaluations, overall accuracy rises from approximately 50% (prior models) to 72.58%, with a density score of 145.16, reflecting improved information preservation and visual coherence. The VLM-guided repair loop effectively reduces layout errors, leading to more visually appealing posters. These results confirm that the integrated approach effectively handles diverse sources, maintains factual fidelity, and produces high-quality visual summaries, setting a new benchmark in multimodal poster generation.
Applications
This technology can be immediately applied in academic conference preparation, enabling researchers to upload their papers or data and receive ready-to-use posters. In corporate settings, it facilitates rapid visualization of financial reports, market analyses, and project summaries, saving time and enhancing presentation quality. Educational platforms can leverage the system to generate visual aids from lecture notes, videos, and web content, improving learning engagement. Long-term, the system could evolve into an intelligent content dissemination platform, automatically transforming raw data into visually compelling narratives for media, marketing, and public outreach, thereby democratizing high-quality visual communication across sectors.
Plain Language Accessible to non-experts
想象你在准备一个学校的展览，你有很多不同的资料，比如科学论文、网页、幻灯片、手写笔记和视频。每个资料都像不同的拼图碎片，内容丰富但格式各异。你需要把这些拼图拼成一张漂亮的海报，让大家一眼就能明白你的主题。传统方法就像用模板拼图，只能拼特定的拼图，不能应对不同的资料。而这项新技术就像有一台神奇的拼图机，它能理解各种不同的碎片，自动整理出一张漂亮的海报。它会分析每个资料的重点，把重要的内容放大，排版得整整齐齐，还能自动修正排版中的问题，比如文字太挤或图片比例不对。最终，你只需要提供资料，这台拼图机就能帮你做出一张专业、漂亮的海报，省时又省力，让你的展览更吸引人。
ELI14 Explained like you're 14
想象你在学校准备一个展示板，你有很多不同的资料，比如科学论文、网页、PPT、笔记和视频。每个资料都像不同的拼图块，有的长，有的短，有的内容很复杂。你想把它们拼成一张漂亮的海报，让别人一看就明白你的主题。以前，你得用模板，把每个资料都按一样的格式排好，非常麻烦。而现在，有了这个新技术，就像有一台超级智能的拼图机器人，它可以理解各种不同的资料，把重要的内容提取出来，然后自动帮你排版，设计出一张漂亮的海报。它还能修正排版中的问题，比如文字太挤或图片比例不对。你只需要把资料交给它，它就能帮你做出一份专业的海报，节省了很多时间，也让你的展示更吸引人。
Abstract
Visual posters are a compact medium for communicating dense information, yet progress on automatic poster generation remains difficult to measure because existing evaluations are often restricted to paper-only inputs, narrow domains, or surface-level visual similarity. We introduce Any2Poster Bench, a benchmark for any-source poster generation that evaluates systems across eight input modalities--PDFs, URLs, PPTX, DOCX, Markdown, LaTeX, notebooks, and videos--and five content domains. Any2Poster Bench pairs each source with quiz-based probes of verbatim factual retention and interpretive understanding, together with VLM-based judgments of visual quality, layout, readability, content completeness, and logical flow, enabling reproducible assessment of both information fidelity and visual communication. To instantiate and validate this benchmark, we further present Any2Poster Agent, an end-to-end reference agent that parses heterogeneous sources, organizes salient content, plans poster layouts, renders posters, and iteratively refines them using visual feedback. On Any2Poster Bench, Any2Poster Agent achieves 87.25% average accuracy across input modalities and 87.28% across content domains. On PaperQuiz-style evaluation, where prior paper-to-poster agents are directly comparable, Any2Poster Agent improves over PosterAgent-4o from 51.06-51.33% to 72.58% overall accuracy and from 116-121 to 145.16 in density-augmented score. Together, Any2Poster Bench and Any2Poster Agent provide a reusable evaluation resource and a competitive baseline for studying multimodal, domain-general poster generation.
References (20)
D2S: Document-to-Slide Generation Via Query-Based Text Summarization
Edward Sun, Yufang Hou, Dakuo Wang et al.
ART: Automatic multi-step reasoning and tool-use for large language models
Bhargavi Paranjape, Scott M. Lundberg, Sameer Singh et al.
Toolformer: Language Models Can Teach Themselves to Use Tools
Timo Schick, Jane Dwivedi-Yu, Roberto Dessì et al.
PAL: Program-aided Language Models
Luyu Gao, Aman Madaan, Shuyan Zhou et al.
ReAct: Synergizing Reasoning and Acting in Language Models
Shunyu Yao, Jeffrey Zhao, Dian Yu et al.
PosterBot: A System for Generating Posters of Scientific Papers with Neural Models
Sheng Xu, Xiaojun Wan
Text2Poster: Laying Out Stylized Texts on Retrieved Images
Chuhao Jin, Hongteng Xu, Ruihua Song et al.
Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI
Mahima Pushkarna, Andrew Zaldivar, Oddur Kjartansson
SlideGen: an abstractive section-based slide generator for scholarly documents
Athar Sefid, P. Mitra, C. Lee Giles
G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment
Yang Liu, Dan Iter, Yichong Xu et al.
PubLayNet: Largest Dataset Ever for Document Layout Analysis
Xu Zhong, Jianbin Tang, Antonio Jimeno-Yepes
GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering
Drew A. Hudson, Christopher D. Manning
Datasheets for datasets
Timnit Gebru, Jamie H. Morgenstern, Briana Vecchione et al.
Learning to Generate Posters of Scientific Papers by Probabilistic Graphical Models
Yuting Qiang, Yanwei Fu, Xiao Yu et al.
VQA: Visual Question Answering
Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol et al.
PosterGen: Aesthetic-Aware Paper-to-Poster Generation via Multi-Agent LLMs
Zhilin Zhang, Xiang Zhang, Jiaqi Wei et al.
Presentations by the Humans and For the Humans: Harnessing LLMs for Generating Persona-Aware Slides from Documents
I. Mondal, S. S, Anandhavelu Natarajan et al.
Design2Code: Benchmarking Multimodal Code Generation for Automated Front-End Engineering
Chenglei Si, Yanzhe Zhang, Ryan Li et al.
PosterSum: A Multimodal Benchmark for Scientific Poster Summarization
Rohit Saxena, Pasquale Minervini, Frank Keller
OctoTools: An Agentic Framework with Extensible Tools for Complex Reasoning
Pan Lu, Bowen Chen, Sheng Liu et al.
Cited By (3)
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs
ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog