PaperlayerView exact revision on arXiv

Source-linked AI summary

Any2Poster: Any-Source Poster Generation Across Modalities and Domains

Amogh Vinaykumar, Aiden Li, Suozhi Huang, Shilong Liu

arXiv:2606.02915v1cs.CV

TL;DR

Automatic poster-generation evaluation is limited by paper-only inputs, narrow domains, and surface-level visual measures. Any2Poster introduces a broad benchmark and reference agent, achieving strong cross-modality and cross-domain performance while improving paper-to-poster results over PosterAgent-4o.

  • Problem

    Existing poster-generation evaluations mainly focus on single input formats, especially academic PDFs, leaving limited evidence for generalization across modalities and domains.

    Source: §1 · pp. 1–2

  • Method

    Any2Poster Bench evaluates eight modalities and five domains with BenchQuiz and VLM-as-judge assessment, while Any2Poster Agent uses unified parsing, adaptive planning, HTML/CSS rendering, and VLM-guided repair.

    Source: §1 · p. 2

  • Results

    87.25% average accuracy across modalities and 87.28% across domains accompany improved PaperQuiz-style performance over PosterAgent-4o.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Any2Poster Bench and Any2Poster Agent provide a reusable evaluation resource and reference baseline for multimodal, domain-general poster generation.

    Source: §6 · p. 9

    The benchmark does not exhaustively cover all poster domains, languages, or visual styles and is not a general-purpose measure of visual design ability independent of information fidelity.

    Source: §3.4 · p. 5

Abstract

from arXiv · show

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.

1 Introduction

Any2Poster addresses the gap between paper-only poster generation and general-purpose generation by benchmarking diverse sources and providing an agent that parses, plans, renders, and refines posters. It achieves strong performance across modalities, domains, and paper-to-poster evaluation.

1 Introduction

  • Any2Poster Bench expands evaluation beyond paper-only inputs to eight modalities and five content domains, measuring cross-modality and cross-domain generalization.The benchmark targets informative, readable, and visually coherent single-page posters.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Any2Poster Agent combines unified parsing, content-adaptive planning, editable HTML/CSS rendering, and VLM-guided localized repair.The shared structured representation supports downstream planning and rendering across heterogeneous source formats.Source: §1 · p. 2
  • 87.25% average accuracy across eight input modalities and 87.28% across five content domains demonstrate Any2Poster Agent’s broad evaluation performance.The agent also improves over GPT-5 by 3.35 points in cross-modality accuracy and over GPT-4o by 11.78 points in cross-domain accuracy.Source: §1 · p. 2
  • 72.58% overall accuracy and 145.16 density-augmented score improve over PosterAgent-4o’s 51.06–51.33% and 116–121, respectively.These results come from the comparable PaperQuiz-style paper-to-poster setting.Source: §1 · p. 2
  • BenchQuiz tests verbatim factual retention and interpretive understanding, complemented by VLM-as-judge assessment of visual and communicative quality.The evaluation therefore covers both information fidelity and visual quality rather than surface-level similarity alone.Source: §1 · p. 2

2 Related Work

Prior work largely targets paper-to-poster or multi-page presentation generation, while Any2Poster combines heterogeneous source understanding, single-page layout reasoning, and agentic visual refinement. Any2Poster Bench extends evaluation to information recovery and visual communication across modalities and domains.

2 Related Work

  • Any2Poster differs from paper-only poster and slide-generation systems by targeting heterogeneous inputs with unified parsing, content-adaptive planning, and editable HTML/CSS rendering.Slides distribute information across pages, whereas posters compress it into one canvas, increasing sensitivity to density, hierarchy, readability, and balance.Source: §2 · p. 3, §2.1 · p. 3

2.2 Document Understanding, Layout, and Rendering

  • Unified parsing is motivated by document-understanding tools that normalize heterogeneous inputs into a shared structured representation for downstream reasoning.The shared representation replaces modality-dependent heuristics with structure-based planning and rendering.Source: §2.2 · p. 3
  • Any-source poster generation must jointly extract source information, summarize it, select or synthesize visuals, and render an editable single-page poster.This combination distinguishes the task from isolated layout, rendering, or document-conversion work.Source: §2.2 · p. 3

2.3 Agentic Generation and Evaluation

  • Agentic generation is well suited to posters because systems must parse inputs, plan layouts, render editable outputs, inspect visual artifacts, and repair failures.This workflow connects language-model reasoning and tool use with iterative visual refinement.Source: §2.3 · p. 3
  • Any2Poster Bench extends PaperQuiz-style evaluation by measuring quiz-based information recovery and VLM-as-judge visual communication quality across modalities and content domains.The benchmark evaluates both what information survives and how effectively the poster communicates it.Source: §2.3 · pp. 3–4, Figure 2 · p. 4

3 Any2Poster Bench

Any2Poster Bench is a controlled benchmark for transforming diverse sources into single-page posters, combining source-grounded comprehension tests with visual-quality judgments. Its scope is deliberately bounded to cross-modality and cross-domain poster evaluation rather than exhaustive design assessment.

3 Any2Poster Bench

  • Any2Poster Bench contains approximately 160 sources spanning eight modalities and five content domains to test generalization beyond paper-specific poster generation.The modalities are PDF, URL, PPTX, DOCX, Markdown, LaTeX, notebook, and video.Source: §3 · p. 4

3.1 Benchmark Construction

  • Benchmark sources are selected for public accessibility, sufficient information density, and diversity in modality and domain.These principles support documented sources, meaningful posters, and non-trivial comprehension evaluation.Source: §3.1 · p. 4

3.2 BenchQuiz Evaluation Protocol

  • BenchQuiz uses 20 verbatim and 20 interpretive multiple-choice questions per source to test factual recovery and higher-level understanding.VLM readers answer using only the generated poster, and raw accuracy is the primary metric.Source: §3.2 · p. 5
  • Answer options are randomly shuffled and VLM readers may answer NA when the poster lacks sufficient evidence, focusing scores on communicated information rather than prior knowledge.Questions are source-grounded and designed to measure information transfer through the poster.Source: §3.2 · p. 5

3.3 VLM-as-Judge Evaluation

  • VLM-as-judge scores element quality, layout balance, engagement, clarity, content completeness, and logical flow, yielding aesthetic, information, and overall scores.The protocol complements BenchQuiz by assessing visual readability and organization.Source: §3.3 · p. 5

3.4 Dataset Documentation and Intended Use

  • The benchmark is not an exhaustive measure of factual correctness, visual design ability, poster domains, languages, or visual styles.It is intended as a controlled testbed for cross-modality and cross-domain generalization in single-page poster generation.Source: §3.4 · p. 5

4 Any2Poster Agent

Any2Poster Agent converts heterogeneous sources into editable posters through unified parsing, content-adaptive planning, HTML/CSS rendering, and localized VLM-guided repair.

4.1 Framework

  • The framework combines unified parsing, content-adaptive planning, and code-based rendering to support diverse modalities and efficient refinement.Its six operational stages are parse, chunk, analyze, plan, generate, and compile.Source: §4.1 · pp. 5–6

4.2 Unified Parsing

  • Unified parsing normalizes PDFs, documents, slides, Markdown, URLs, videos, LaTeX, and notebooks into one structured representation.The representation includes sections, figures with captions, tables, and raw text, enabling downstream reasoning over content structure rather than modality-specific heuristics.Source: §4.2 · p. 6

4.3 Content-adaptive Poster Planning

  • Content-adaptive planning assigns panel roles, layout modes, visual treatments, and source-versus-generated visuals according to content and domain.Global and section-level passes produce poster-ready content, while weakly structured inputs receive narrative segmentation before panelization.Source: §4.3 · p. 7

4.4 HTML/CSS-based Rendering

  • HTML/CSS rendering keeps poster text editable and separates typography, layout, and visual assets for fine-grained control.Generated visuals use panel concepts, visual types, validated data, and the active design theme; unsupported data values are discarded before generation.Source: §4.4 · p. 7

4.5 Visual Feedback Loop with VLMs

  • VLM-guided panel repair detects overflow, blank space, and scaling problems, then applies targeted HTML/CSS edits without full-poster regeneration.Localized repair preserves correct content while improving readability and layout quality.Source: §4.5 · p. 7

5 Experiments

Experiments evaluate information recovery and visual communication across modalities, domains, prior paper-to-poster systems, and component ablations. Any2Poster Agent performs strongly across settings, with gains in broad accuracy, visual organization, and interpretive understanding.

5 Experiments

  • The evaluation combines verbatim and interpretive quiz accuracy with VLM judgments of element quality, layout, engagement, clarity, completeness, and logical flow.The experiments test modality generalization, domain generalization, prior-agent competitiveness, and preservation of recoverable information.Source: §5 · p. 7, §5.1 · p. 8

5.2 Cross-Modality Evaluation

  • 87.25% average accuracy across eight modalities makes Any2Poster Agent the strongest cross-modality system, although PPTX reaches 75.00%.Markdown and LaTeX exceed 93%, while GPT-5 is slightly better on PDF and PPTX.Source: §5.2 · p. 8, Table 2 · p. 8

5.3 Cross-Domain Evaluation

  • 87.28% average accuracy across five domains exceeds GPT-4o’s 75.50% and Paper2Poster’s 68.50% on supported domains.All domain-level accuracies for Any2Poster Agent exceed 86%.Source: §5.3 · p. 8, Table 3 · p. 8

5.4 Comparison with Prior Paper-to-Poster Agents

  • 72.58% overall accuracy and 145.16 density-augmented score improve over PosterAgent-4o’s 51.06–51.33% and 116–121.Interpretive accuracy rises from 49.17% to 84.00%, indicating stronger communication of high-level meaning.Source: §5.4 · p. 9, Table 4 · p. 9

5.5 VLM-as-Judge Results

  • 4.03 overall VLM-as-judge score improves over Paper2Poster’s 3.69, with logical flow increasing from 3.64 to 4.96.Element quality, layout balance, engagement, clarity, content completeness, aesthetic score, and information score also improve.Source: §5.5 · p. 9, Table 5 · p. 9

5.6 Ablation Study

  • Removing unified parsing reduces verbatim accuracy from 60.80% to 31.00%, while removing visual feedback reduces overall accuracy from 72.58% to 55.50%.These ablations associate structure-aware parsing with fine-grained fact preservation and visual refinement with recoverable information.Source: §5.6 · p. 9

6 Conclusion

The paper introduces Any2Poster Bench for evaluating any-source poster generation across modalities and domains, and Any2Poster Agent as a reference system for producing and refining such posters.

6 Conclusion

  • Any2Poster Bench measures information recoverability and visual communication across eight input modalities and five content domains.It combines BenchQuiz with VLM-as-judge assessment.Source: §6 · p. 9
  • Any2Poster Agent combines unified parsing, content-adaptive planning, HTML/CSS rendering, and VLM-guided repair.Experiments report generalization across modalities and domains and improvement over prior paper-to-poster agents.Source: §6 · p. 9

A Dataset Details

The benchmark and prompt suite define heterogeneous source coverage, structured processing, poster construction, feedback, and quiz- and VLM-based evaluation procedures.

A.1 Benchmark Composition

  • The benchmark spans approximately 160 sources across eight modalities and five domains, with 32 fully evaluated instances released.The remaining instances are withheld when source licenses prohibit redistribution of extracted text.Source: §A.1 · p. 14
  • Sources are selected for accessibility, information density, and heterogeneous structure beyond scientific papers.Included structures range from papers and slide decks to webpages, notebooks, video transcripts, and weakly structured prose.Source: §A.1 · p. 14

A.2 Source Processing

  • Format-specific parsers convert documents, URLs, videos, notebooks, and LaTeX projects into document-like structured representations.Processing extracts sections, figures, captions, tables, raw text, transcripts, code, outputs, and associated assets when available.Source: §A.2 · p. 14

B.1 Analysis Prompts

  • The prompt suite uses global analysis and per-section distillation to produce poster metadata, content, section structure, and visual suggestions.Headerless documents receive prose segmentation, while prompts constrain section count, distinct coverage, and headline-result emphasis.Source: §B.1 · pp. 14–15

B.2 Generation Prompts

  • Generation prompts specify compact readable HTML/CSS panels, clear typography, safe padding, and content-specific vector visuals.Visual prompts require readable labels, high contrast, background matching, and specificity to the source content.Source: §B.2 · p. 16

B.3 Feedback Prompts

  • A feedback loop inspects panel crops and applies deterministic HTML/CSS mutations to address detected layout issues.The evaluation prompts support PaperQuiz comparability, BenchQuiz for non-research documents, and VLM-as-judge scoring.Source: §B.3 · p. 16, §B.4 · p. 16

B.4 Evaluation Prompts

  • PaperQuiz prompts separately test verbatim retrieval and interpretive understanding, with poster-only answering and NA when evidence is insufficient.VLM-as-judge prompts score visual elements and layout using conservative 1–5 criteria.Source: §B.4.1 · pp. 16–17, §B.4.2 · p. 17, §B.4.4 · p. 17

B.4.3 BenchQuiz: Genre-Aware Evaluation

  • BenchQuiz uses 20 genre-aware questions per non-research document, split between concrete-detail retention and high-level interpretation.Questions are answered from the poster and test explicit facts, central arguments, significance, and takeaways.Source: §B.4.3 · p. 17

C Implementation Details

Any2Poster Agent converts heterogeneous sources into structured poster plans, renders them as editable HTML/CSS, and applies localized VLM-guided repairs to visual issues.

C.1 Any2Poster Agent

  • The agent parses heterogeneous inputs into a shared schema, chunks long sections, analyzes global and section-level content, and plans panel roles, layouts, and visuals.Its analysis identifies the title, core message, section importance, candidate panels, source figures, summaries, layout modes, and visual suggestions.Source: §C.1 · p. 18
  • Posters are rendered in HTML/CSS, with direct HTML text for character-level fidelity, extracted or generated visuals, and PDF/PNG export.A VLM can crop panels, diagnose visual issues, and apply localized HTML/CSS edits before final rendering.Source: §C.1 · p. 18

C.2 Visual Feedback Rules

  • The feedback loop maps panel diagnoses to deterministic repairs: overflow shrinks text, blank panels increase visual prominence or spacing, and figure sizing is adjusted locally.Local repair avoids regenerating the full poster and preserves content that is already correct.Source: §C.2 · p. 18

D Ablation Study

The ablation study evaluates Any2Poster Agent components in the PaperQuiz-style setting and finds that unified parsing and visual feedback materially support information recovery.

D Ablation Study

  • Removing unified input parsing causes the largest drop in verbatim accuracy, highlighting the importance of structured source recovery for fine-grained facts.The result indicates that parsing quality is especially important for retaining exact information from heterogeneous sources.Source: D Ablation Study · p. 18
  • Removing visual feedback refinement also reduces overall accuracy, showing that visual quality control affects whether readers can recover information from the final poster.The ablation uses the PaperQuiz-style setting with 50 verbatim and 50 interpretive questions.Source: D Ablation Study · p. 18, Table 6 · p. 18

E Efficiency

Any2Poster Agent-4o uses moderate computational resources while supporting a full heterogeneous-input poster pipeline, and qualitative examples illustrate coverage across source structures and domains.

E Efficiency

  • 138.34K tokens per run is the average consumption of Any2Poster Agent-4o, higher than PosterAgent-4o and direct 4o-HTML but lower than PPTAgent-4o and OWL-4o.The comparison characterizes computational cost across poster-generation methods.Source: E Efficiency · p. 18, Table 7 · p. 19
  • The direct 4o-HTML baseline has the lowest token consumption, but it represents simpler direct generation rather than a full agentic pipeline.Any2Poster Agent-4o therefore has moderate cost while supporting parsing, planning, generation, compilation, and feedback refinement.Source: E Efficiency · p. 19

F Additional Qualitative Examples

  • Additional qualitative examples cover technical papers, webpages, notebooks, slide decks, and narrative prose across the benchmark’s heterogeneous sources.Figure 4 compares generated posters from different systems on the same source.Source: F Additional Qualitative Examples · p. 19, Figure 4 · p. 19

G Limitations

Any2Poster Bench broadens evaluation beyond paper-only inputs but remains bounded by its current languages, modalities, domains, evaluation models, and reference-agent capabilities.

G Limitations

  • The benchmark currently focuses on English-language sources and eight modalities and five domains, leaving broader languages, specialized domains, and additional source types for future versions.Suggested extensions include interactive webpages, spreadsheets, and multi-file project repositories.Source: G Limitations · p. 19
  • BenchQuiz and VLM-as-judge scores may not fully match human behavior because reader performance and visual-quality judgments can reflect model-specific preferences.Answer-option randomization and NA responses reduce bias, but residual model bias may remain.Source: G Limitations · pp. 19–20
  • Any2Poster Agent is a reference system whose performance depends on parser, planning, visual-generation, and feedback-loop quality.Heavily formatted slides, noisy videos, complex tables, precise visual reproduction, and large-model computational costs remain practical challenges.Source: G Limitations · p. 20

H Responsible Release

  • The benchmark should not be used to claim factual correctness beyond source content or as a standalone measure of general design ability.Release materials document provenance, licensing, intended uses, limitations, and recommended evaluation practices, with metadata or pointers used when redistribution is restricted.Source: H Responsible Release · p. 20

Related papers

Loading 2606.02915v1…