Learning to Generate Posters of Scientific Papers by Probabilistic Graphical Models

Motivation

• Generating high-quality scientific posters requires jointly understanding paper content and applying design expertise, making the task difficult and time-consuming.Posters must remain readable, informative, and visually aesthetic, particularly for researchers unfamiliar with design tools.Source: §1 · p. 1

Problem

• Scientific poster generation requires extracting content, fitting it into readable panel layouts, and carefully arranging graphical elements.These challenges explain why the paper identifies no existing automatic tools for scientific poster generation.Source: §1 · p. 1

Approach

• The proposed method uses TextRank and graphical-content extraction, then infers panel attributes, recursively generates panel layouts, and models graphical-element composition with probabilistic graphical models.The framework targets layout generation while incorporating paper content and graphical elements.Source: §1 · pp. 1–2

Results

• Compared with author-designed posters, the approach is more efficient and versatile and can adapt to different paper sizes, aspect ratios, and styles.Adaptation is achieved by training the model with different datasets.Source: §1 · p. 2

Contributions

• The paper presents the first method for generating scientific posters from academic papers and extends prior work with an enlarged dataset, improved losses and modeling, and additional experiments.The released benchmark dataset supports evaluation of the task.Source: §1 · p. 2

2.1 General Graphical Design

• General graphical-design methods prioritize plausible visual arrangements and aesthetics, whereas scientific posters must also preserve textual content, paper structure, and content order.Existing examples include photo albums, furniture arrangements, interfaces, circuit boards, and graph layouts.Source: §2.1 · p. 2

2.2 Comic Layout Generation

• Comic-layout methods arrange panels and can model text balloons or reader attention, but they do not address combined text and graphical-element composition within scientific-poster panels.The paper’s panel-layout method instead uses a binary tree and recursively searches splitting configurations.Source: §2.2 · p. 2

2.3 Presentation Layout Automation

• Presentation-layout automation commonly focuses on micro-typography, templates, or page-layout selection, while this paper addresses both panel layout and graphical-element sizing.Unlike prior rule-based approaches, it learns layout patterns from annotated training data.Source: §2.3 · p. 2
• Other single-page and aesthetics-driven methods treat text as rectangular blocks or omit constraints on content ordering, which is unsuitable for scientific posters.Scientific-poster generation must retain text flow and the ordering of paper content.Source: §2.3 · pp. 2–3

Task Definition

• A scientific poster maps paper sections to panels containing explanatory text, bullet points, and important figures or tables.This structure is intended to preserve the paper’s key messages while supporting readability, informativeness, and visual aesthetics.Source: §3 · p. 3

Dataset

• The NJU-Fudan Paper-Poster dataset contains 85 computer-science paper-poster pairs collected from an online website.The dataset was introduced to further study scientific-poster generation.Source: §3 · p. 3

Annotations

• The dataset annotates poster layout attributes such as panel positions and figure sizes, alongside content attributes such as text length in each panel.Related layout information, including figure size in the original paper, is also manually labelled.Source: §3 · p. 3

Resources

• The authors provide an annotation tool and release the dataset and tooling to support annotation and labelling of additional data.The released resources are intended to facilitate further research on the task.Source: §3 · p. 3

Problem Formulation

• Each poster consists of panels containing graphical elements, with panel attributes covering text, graphical content, size, and aspect ratio.Panel attributes include text length and ratio, graphical-element count and ratio, panel size, and aspect ratio.Source: §4 · p. 3
• Graphical elements are characterized by their size and aspect ratio, while poster layout also includes relative horizontal position and graphical-element size within a panel.These attributes describe both panel-level structure and within-panel composition.Source: §4 · pp. 3–4

Overview

• The framework aims to determine these attributes for every panel and graphical element together with the arrangement of panels.The goal is to learn a coordinated poster layout rather than isolated attribute values.Source: §4 · p. 4
• A Bayesian network models relationships among paper-derived and predicted attributes, then uses likelihood-weighted sampling to infer layouts.This is preferred to treating all features as a single regression input because it exposes relationships such as those between panel size and graphical content.Source: §4 · p. 4

5.1 Content Extraction

• TextRank summarizes each paper section, while user interaction ranks important figures and tables for graphical-content extraction.Source: §5.1 · p. 4

5.2 Panel Attributes Inference

• A Bayesian network infers each panel’s initial size and aspect ratio from text ratio, graphical-element count, and graphical-element ratio.The model uses conditional probability distributions estimated by maximum likelihood, assuming conditional independence of size and aspect ratio and independence among panel attribute sets.Source: §5.2 · p. 4

5.3 Panel Layout Generation

• Panel layouts use a binary tree and recursive page splitting to update panel attributes while balancing shape variation, relative panel sizes, and partition symmetry.The algorithm compares candidate horizontal and vertical splits and selects arrangements with lower combined loss, including a penalty for asymmetric splitting ratios.Source: §5.3 · pp. 4–5
• Figure 3 illustrates a five-panel tree whose first vertical split creates two equal columns, followed by recursively chosen horizontal subdivisions.The left column contains three panels using a 0.33/0.67 split and a subsequent 0.5/0.5 split; the right column is split once at 0.5/0.5.Source: Figure 3 · p. 5

5.4 Composition within a Panel

• Within each panel, a Bayesian network samples graphical-element positions and sizes from panel and element attributes before filling textual content sequentially.Horizontal position depends on panel shape and element properties, while element size depends on panel attributes and the element’s own properties.Source: §5.4 · pp. 5–6
• Model parameters for panel composition are learned from observed training data by maximizing complete-data likelihood, with closed-form estimates for conditional Gaussian parameters and IRLS for soft-max parameters.The conditional Gaussian parameters are estimated directly, whereas soft-max parameters require iterative reweighted least squares.Source: §5.4 · p. 6
• MAP-based panel composition combines Bayesian-network likelihood with a term favoring content that fills panel size precisely, using likelihood-weighted sampling for approximate inference.Exact MAP inference is intractable in the model, so the method uses likelihood-weighted sampling to incorporate the design constraint.Source: §5.4 · p. 6

6.1 Experimental Setup

• The framework is faster than manual poster design and novice-generated posters, according to the reported running-time evaluation.Experiments ran on a PC with an Intel Xeon 3.6 GHz CPU and 11.6 GB RAM.Source: §6.1 · p. 7, Table 1 · p. 7

6.2 Quantitative Evaluation

• RMSE for panel size and aspect ratio is 0.71 and 0.695, respectively, lower than all compared methods.The model uses tp and gp as features to infer these panel attributes.Source: §6.2 · p. 7, Table 2 · p. 7
• The method beats all other methods for graphical-element size and horizontal-position prediction when evaluated with RMSE and accuracy.The authors attribute this result to design constraints incorporated during inference.Source: §6.2 · p. 7, Table 3 · p. 7

6.3 Qualitative User Study Evaluation

• In a user study, generated posters were comparable to original posters in readability and informativeness and significantly better than novice posters, using identical contents across variants.Researchers scored three results shown in randomized order from 0–10, and final scores were averaged across subjects.Source: §6.3 · p. 8
• The generated posters score below original posters on aesthetics but above novice posters, reflecting the subjectivity and human interaction involved in aesthetic refinement.The paper identifies aesthetic generation as an open problem.Source: Table 4 · p. 7, §6.3 · p. 8
• The generated panel layouts resemble original layouts more closely than novice layouts, while graphical-element sizes appear similar to those in original posters.The authors relate this to the dataset’s consistent design style and their recursive split and panel-layout algorithm.Source: §6.3 · p. 8

6.4 Qualitative Evaluation by Design Principles

• Recursive page splitting preserves section order in the panel read-order, while sequential within-panel organization supports readability.The method also avoids significant panel overlap and boundary crossings through recursive splitting, sequential arrangement, and design constraints.Source: §6.4 · p. 8

• Learning design patterns from existing examples is proposed as an assistant tool for scientific poster generation.The approach is intended to help designers with the difficult task of balancing utility and aesthetics.Source: §7 · p. 8
• The framework may extend to web-page and single-page graphical design when corresponding layout styles are available.Source: §7 · p. 8
• Font types are not considered in the current framework and are identified for future work.Source: §7 · p. 8