PaperlayerView exact revision on arXiv

Source-linked AI summary

PosterGen: Aesthetic-Aware Multi-Modal Paper-to-Poster Generation via Multi-Agent LLMs

Zhilin Zhang, Xiang Zhang, Jiaqi Wei, Yiwei Xu, Chenyu You

arXiv:2508.17188v2cs.AI

TL;DR

Academic poster generation is time-consuming, and existing approaches often neglect design principles, requiring substantial manual refinement. PosterGen uses specialized agents to extract and organize paper content, construct layouts, apply styling, and render posters. It achieves content and aesthetic quality comparable to human-designed posters while outperforming existing multi-agent methods in visual design.

  • Problem

    Existing academic poster-generation approaches often neglect layout, color, typography, and other design principles, leaving researchers with substantial manual refinement.

    Source: §1 · p. 1

  • Method

    PosterGen uses collaborative specialized agents for content distillation, ABT storyboard construction, spatial layout, cohesive styling, and poster rendering.

    Source: §1 · p. 2, §4.3 · p. 4

  • Results

    PosterGen achieves content and aesthetic quality comparable to human-designed posters and consistently outperforms state-of-the-art multi-agent methods across almost all metrics.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    PosterGen produces visually compelling, presentation-ready academic posters while substantially alleviating researchers’ manual preparation burden.

    Source: §1 · p. 2, §6 · p. 8

    ParserAgent can occasionally fail to extract text or low-resolution visual components, while CuratorAgent is limited to source-paper visuals and LayoutAgent needs more diverse layouts.

    Source: F. Limitations · p. 15

Abstract

from arXiv · show

Multi-agent systems built upon large language models (LLMs) have demonstrated remarkable capabilities in tackling complex compositional tasks. In this work, we apply this paradigm to the paper-to-poster generation problem, a practical yet time-consuming process faced by researchers preparing for conferences. While recent approaches have attempted to automate this task, most neglect core design and aesthetic principles, resulting in posters that require substantial manual refinement. To address these design limitations, we propose PosterGen, a multi-agent framework that mirrors the workflow of professional poster designers. It consists of four collaborative specialized agents: (1) Parser and Curator agents extract content from the paper and organize storyboard; (2) Layout agent maps the content into a coherent spatial layout; (3) Stylist agents apply visual design elements such as color and typography; and (4) Renderer composes the final poster. Together, these agents produce posters that are both semantically grounded and visually appealing. To evaluate design quality, we introduce a vision-language model (VLM)-based rubric that measures layout balance, readability, and aesthetic coherence. Experimental results show that PosterGen consistently matches in content fidelity, and significantly outperforms existing methods in visual designs, generating posters that are presentation-ready with minimal human refinements.

1. Introduction

Academic poster generation is time-consuming, while existing automated methods often neglect design principles and require manual refinement. PosterGen addresses this gap with an aesthetic-aware multi-agent workflow and evaluation showing strong content and visual quality.

  • PosterGen targets the manual burden of designing high-quality academic posters, a process needed for conference communication.Source: §1 · p. 1
  • Existing academic poster methods address partial tasks or produce content and design problems, including overflow, weak reading flow, and limited styling.Source: §1 · p. 1
  • PosterGen integrates content distillation, ABT storyboard construction, balanced layout, and cohesive color and typography through specialized agents.Source: §1 · p. 2, Figure 1 · p. 2
  • PosterGen is evaluated with VLM-as-Judge and human studies, measuring content fidelity and aesthetic quality against existing methods.Source: §1 · p. 2
  • PosterGen achieves content and aesthetic quality comparable to human-designed posters while outperforming state-of-the-art multi-agent methods across almost all metrics.Source: §1 · p. 2

2. Related Work

Prior work spans artistic poster generation, academic poster sub-tasks, and LLM-based multi-agent systems, but academic poster methods remain constrained in content fidelity and aesthetic design.

  • Artistic and commercial poster systems emphasize visual appeal, control, or rendering, but academic posters additionally require precise communication of complex research.Source: §2 · pp. 2–3
  • Academic poster research includes neural generation, layout generation, and text summarization, but these approaches can suffer content overflow or narrow task coverage.Source: §2 · p. 2
  • P2P [44] and PosterAgent [32] introduced multi-agent approaches for scientific posters, yet they lack thorough design consideration and require extensive manual adjustment.Source: §2 · p. 2
  • Figure 2 contrasts chaotic reading flow from vertically unaligned layouts with the natural flow and eye-level visual emphasis of vertically aligned grids.Source: Figure 2 · p. 3

3. Design Principles

PosterGen embeds established poster-design principles into its agents, covering narrative, layout structure, color, and typography to support clarity, hierarchy, and readability.

Narrative

  • The ABT structure organizes context, problems, and solutions, guiding content-driven section titles and a coherent poster narrative.Source: §3 · p. 3

Layout Structure

  • A three-column grid supports natural reading flow, positions a key visual at the eye-level hot zone, and uses white space to reduce clutter.Source: §3 · p. 3

Color Design

  • A restrained monochromatic palette establishes hierarchy through theme colors, background variants, and high-contrast accents while meeting a WCAG 4.5:1 text contrast ratio.Source: §3 · p. 3

Typography Design

  • Typography reinforces clarity through legible sans-serif typefaces, font-size hierarchy, and semantic emphasis using formatting and contrast color.Source: §3 · p. 3

4. Method

PosterGen uses a cascade of specialized agents to transform a research paper into a structured, aesthetically designed PPTX poster. The workflow combines narrative organization, spatial layout, styling, and rendering, with explicit mechanisms for content extraction, whitespace control, color harmony, typography, and evaluation.

4. Method

  • PosterGen combines Parser, Curator, Layout, Color, Font, and Renderer components to impose structured design constraints throughout poster generation.The workflow extracts and organizes paper content, constructs a storyboard, maps it to a layout, applies colors and fonts, and renders the final file.Source: §4 · p. 3

4.1. ParserAgent

  • ParserAgent converts a PDF into Markdown and visual assets, then distills the paper into ABT narrative sections and classifies visuals by narrative role.The processing includes a maximum of 1000 words per section and categories such as key visual, problem illustration, workflow, main results, comparative results, and supporting material.Source: §4.1 · pp. 3–4

4.2. CuratorAgent

  • CuratorAgent maps parsed content onto a narrative-driven three-column storyboard while limiting the poster to five to eight sections to reduce overflow.Its spatial narrative design keeps content centered on the ABT structure and minimizes later content and visual refinements.Source: §4.2 · p. 4

4.3. LayoutAgent

  • LayoutAgent transforms the storyboard into a precise three-column layout using procedural geometry, an LLM-based Balancer, height estimation, and a CSS-like box model.Binary search estimates text-frame height, while margin and padding settings provide fine-grained whitespace control for XML-based layouts.Source: §4.3 · p. 4

4.4. StylistAgents

  • StylistAgents treat color and typography as semantic hierarchy tools, generating a cohesive palette and highlighting important section keywords.ColorAgent derives a theme from an affiliation logo or key figure, then creates monochromatic and contrast colors; FontAgent assigns font families, sizes, bolding, and italics.Source: §4.4 · pp. 4–5, Figure 4 · p. 5

4.5. Renderer

  • Renderer converts styled layout metadata into a PPTX, attaches affiliation and conference logos, and produces a PNG through headless LibreOffice for inspection and evaluation.The evaluation criteria include content and design measures on a 1–5 scale, while the content-results table averages scores across 30 posters.Source: §4.5 · p. 5, Table 2 · p. 5, Table 3 · p. 5

5. Experiments

PosterGen is evaluated with VLM and human criteria spanning content fidelity, narrative, layout, color, typography, and technical quality. Across quantitative, qualitative, ablation, and efficiency studies, it generally matches human-designed quality, outperforms baselines, and produces posters in minutes.

5.1. Metrics

  • The evaluation rubric separates Content from Aesthetics and measures information, narrative, layout, color, typography, alignment, and robustness.GPT-4.1 and Claude Sonnet 4 serve as VLM judges to reduce assessment bias.Source: §5.1 · pp. 5–6

5.3. Quantitative Results and Comparisons

  • PosterGen achieves the best or second-best performance across almost every metric, with 4.33 content and 4.43 aesthetic scores from GPT-4.1, comparable to human-designed posters.These results are averaged over 30 posters; the reported comparisons are supported by Tables 3 and 4.Source: §5.3 · p. 6, Table 4 · p. 6

5.4. Human Evaluation

  • Human evaluators rate PosterGen near human-made posters at 4.15 layout, 4.20 color, and 4.17 typography, while outperforming PosterAgent across all aesthetic dimensions.PosterAgent scores about 15% lower in human evaluation than in VLM judging, suggesting greater manual-refinement needs.Source: Figure 5 · p. 6, §5.4 · p. 7

5.2. Baselines

  • PosterGen achieves the highest average content score, 4.33 with GPT-4.1, surpassing P2P [44] and PosterAgent [32] by 13.4%, and the highest average aesthetic score, 4.43.Its aesthetic score is comparable to the human-designed score of 4.27 and outperforms P2P by 11.0%.Source: §5.3 · p. 6

5.5. Qualitative Results

  • Qualitative comparisons show PosterGen approaching human design through title hierarchy, colored section blocks, whitespace, contrasting highlights, and a light monochromatic anchor background.GPT-4o-Image frequently produces gibberish or duplicated content, whereas PosterAgent often has misaligned sections and monotonous styling.Source: §5.5 · pp. 7–8, Figure 6 · p. 7

5.6. Ablation Study

  • The full pipeline obtains the highest content and aesthetic performance, with CuratorAgent preserving information and narrative, LayoutAgent improving spatial quality, and StylistAgents enhancing color and typography.The progressive visualization shows LayoutAgent resolving overflow and column imbalance before stylistic refinement.Source: §5.6 · p. 8, Figure 7 · p. 8

5.7. Efficiency Analysis

  • PosterGen generates a complete poster in 5.32 minutes using eight API calls at $0.20 per poster, while ParserAgent consumes 62.73% of runtime.The remaining agents collectively finish in under two minutes without GPU acceleration or parallelism.Source: §5.7 · p. 8, Table 6 · p. 8

6. Conclusion

PosterGen presents an aesthetic-aware multi-agent framework for academic poster generation that approaches human-level quality. Its design principles, specialized collaboration, VLM evaluation, and user study together support improved aesthetics and reduced manual preparation burden.

6. Conclusion

  • PosterGen is presented as the first aesthetic-aware multi-agent framework to produce visually compelling academic posters approaching human-level quality.The framework mirrors professional designers through specialized agent collaboration and evaluates information fidelity and aesthetic design with VLM-based criteria and a user study.Source: §6 · p. 8

Multi-Agent LLMs

Poster and slide generation both use document-derived content and design automation, but posters must convey a paper’s complete message on one visually appealing page. PosterGen’s agents combine procedural processing with targeted LLM calls, while rendering remains procedural.

A. Extended Related Work

  • Academic posters are harder than slides because they must fit all necessary paper information onto one page without relying on oral explanation.Slides can distribute content across pages and work with a presenter’s narration.Source: A. Extended Related Work · p. 12

B.1. Agent Specifications

  • The implementation uses gpt-4.1-2025-04-14 by default, with content constraints controlled through agent-prompt instructions.The implementation table identifies this as a key hyperparameter.Source: Table 8 · p. 12
  • PosterGen’s agent pipeline mixes single-call LLM agents with hybrid ParserAgent and LayoutAgent components, while Renderer is entirely procedural.CuratorAgent, ColorAgent, and FontAgent use single LLM calls; ParserAgent and LayoutAgent mainly use procedural logic.Source: §B.1 · p. 12, Table 7 · p. 12

B.2. Experimental Environment

The supplementary material documents the software environment, reproducibility parameters, baseline prompt, PosterGen agent prompts, and calibrated VLM-as-Judge evaluation template.

B.2. Experimental Environment

  • PosterGen runs across Windows, Linux, and macOS using Python 3.11, with python-pptx, LangChain, LangGraph, marker-pdf, and Pillow as key dependencies.A GPU is recommended for marker-pdf document parsing and OCR, and marker-pdf is modified to extract tables as images.Source: §B.2 · pp. 12–13

B.3. Configuration Parameters

  • Critical behavior-affecting parameters are listed in Table 8, while aesthetic settings for layout, color, and typography remain in configuration files.The omitted settings include font sizes, margins, and color values.Source: §B.3 · p. 13

C.1. Baseline Prompt

  • The GPT-4o Image baseline receives a prompt through the ChatGPT web interface alongside the input paper file.Source: §C.1 · p. 13

C.2. PosterGen Prompts

  • PosterGen prompts decompose paper understanding into title-and-author extraction, ABT narrative extraction, visual-asset classification, structured-section extraction, and storyboard planning.CuratorAgent additionally selects visual assets, organizes content, and applies visual-height constraints; Layout Balancer improves column utilization and prevents overflow.Source: §C.2 · p. 13
  • ColorAgent extracts theme colors from affiliation logos, while FontAgent uses one LLM call to classify keywords for typography.The ColorAgent prompt also notes a fallback based on a key visual asset.Source: §C.2 · p. 13

C.3. VLM-as-Judge Evaluation Prompt Template

  • The VLM-as-Judge template applies a 5-point scale with positive few-shot examples for scores 4–5 and negative examples for scores 1–3.This design targets overly generous VLM scoring and aims to produce more accurate, evidence-based judgments.Source: §C.3 · p. 13

D. Human Evaluation Details

The human study uses 30 research-experienced participants and blinded randomized comparisons among PosterGen, PosterAgent, and author-designed posters, excluding GPT-4o-Image.

D. Human Evaluation Details

  • 30 participants with research and academic-poster experience evaluated posters through Qualtrics.Participant background and poster-making experience were collected to verify evaluator suitability.Source: D. Human Evaluation Details · p. 13
  • Each participant rated a random subset of 10 poster sets, each containing PosterGen, PosterAgent [32], and author-designed versions.The three posters and the ten sets were independently randomized to reduce ordering bias and maintain blind evaluation.Source: D. Human Evaluation Details · p. 13
  • GPT-4o-Image was omitted because the authors report gibberish text, distorted text blocks, and hallucinated visual assets in qualitative analysis.The authors therefore did not consider it suitable for comparison with multi-agent and human-designed posters.Source: D. Human Evaluation Details · p. 13

E. Additional Qualitative Results

Additional results from a 30-paper benchmark reinforce that PosterGen provides stronger aesthetic control than GPT-4o-Image and PosterAgent while preserving presentation-oriented visual quality.

E. Additional Qualitative Results

  • The supplementary qualitative benchmark covers 15 representative results from 30 papers selected from recent top-tier AI conferences.Included papers had both publicly available full papers and high-quality human-made posters.Source: E. Additional Qualitative Results · p. 13
  • The end-to-end GPT-4o-Image method is prone to basic layout constraints, manifesting serious boundary problems in the supplementary qualitative results.Source: E. Additional Qualitative Results · p. 13
  • PosterAgent [32] improves content fidelity over diffusion-based methods but retains alignment, whitespace, and typographic-aesthetic weaknesses.Source: E. Additional Qualitative Results · p. 14
  • ParserAgent’s ABT extraction prompt requests concise, poster-ready narratives that emphasize key contributions and results while avoiding unnecessary technical jargon.Source: E. Additional Qualitative Results · p. 15
  • PosterGen produces pleasant colors, effective visual hierarchy, and natural reading paths in contrast with the baseline shortcomings.The qualitative evidence attributes these outcomes to incorporating aesthetic concepts into the agent workflow.Source: E. Additional Qualitative Results · p. 15

F. Limitations

PosterGen remains constrained by occasional PDF-extraction errors and a fixed three-column layout, motivating more robust parsing and flexible layout designs.

F. Limitations

  • marker-pdf occasionally fails to extract text or low-resolution visual components, including title or author information.The authors suggest a specialized parser for academic PDFs with intricate formats.Source: F. Limitations · p. 15
  • PosterGen’s three-column grid is a safe landscape-poster choice but limits design flexibility, especially for portrait-oriented and diverse layouts.Future LayoutAgent work should explore varied aspect ratios and dynamic layout designs.Source: F. Limitations · p. 15

G. Broader Impacts

PosterGen embeds professional design principles into an academic-poster generation workflow, aiming to reduce researchers’ manual effort while improving visual quality and preserving content fidelity.

  • PosterGen integrates design and aesthetic principles into a multi-agent academic-poster framework, addressing limitations often overlooked in automated generation.Source: G. Broader Impacts · p. 15
  • PosterGen is intended to relieve poster-creation effort and help researchers focus on the one-on-one scholarly dialogue enabled by poster sessions.Source: G. Broader Impacts · p. 15
  • Visual-asset selection assigns research roles and column placement, including key visuals, problem illustrations, method workflows, and primary or comparative results.Source: G. Broader Impacts · p. 16
  • Prompted agents also support structured section extraction, color selection, keyword-based typography, and evidence-based VLM evaluation across poster-design dimensions.Source: G. Broader Impacts · pp. 17, 23, 24, 25
  • Curator rules exclude visuals exceeding 50% column height, except when a single oversized visual is selected as a fallback.Source: G. Broader Impacts · p. 19

Related papers

Loading 2508.17188v2…