EfficientPosterGen: Semantic-aware Efficient Poster Generation via Token Compression and Accurate Violation Detection

EfficientPosterGen is an end-to-end framework that automates academic poster generation by integrating semantic-aware retrieval, visual-based context compression to reduce token usage, and a deterministic algorithm for reliable layout violation detection, thereby achieving high-quality, token-efficient, and layout-accurate results.

The Gist

Imagine you have written a massive, 50-page research paper. It's full of brilliant ideas, complex data, and detailed arguments. Now, you need to present this at a conference, but you only have a single poster board to work with.

The Problem:
Trying to squeeze that 50-page book onto a single poster is like trying to fit an elephant into a Mini Cooper.

• Too much stuff: If you just copy-paste the whole paper, the text becomes microscopic and unreadable.
• Too expensive: Using a super-smart AI (called a Multimodal Large Language Model) to do this is like hiring a team of 100 architects to design a single shed. It costs a fortune in "tokens" (the currency AI uses to think) and takes forever.
• Messy results: The AI often gets confused, spilling text off the edges of the poster or leaving huge empty white spaces, making it look unprofessional.

The Solution: EfficientPosterGen
The authors of this paper built a new system called EfficientPosterGen. Think of it as a super-efficient, smart editor that doesn't just shrink the paper; it reimagines it. Here is how it works, broken down into three simple steps:

1. The "Highlighter" (Semantic-aware Key Information Retrieval)

Imagine you are reading a novel and need to summarize it for a friend. You wouldn't read every single word; you'd skip the boring descriptions of the weather and focus on the plot twists.

• What it does: This module scans the entire research paper and builds a "map" of how ideas connect. It identifies the "star players" (the most important findings) and ignores the "benchwarmers" (repetitive details, long lists of references, and fluff).
• The Analogy: It's like a curator at a museum. Instead of showing you every single artifact in the vault, they select only the top 10 masterpieces that tell the story of the exhibit.

2. The "Translator" (Visual-based Context Compression)

Usually, you feed text to an AI, and the AI reads it word by word. This is slow and expensive.

• What it does: Instead of sending the AI thousands of words, this module takes the selected text and turns it into images (like screenshots of the text). It then asks the AI to "look" at the image and write the summary.
• The Analogy: Imagine you need to tell a friend a long story.
  • Old way: You read the whole story to them over the phone (takes a long time, costs a lot of minutes).
  • New way: You send them a comic strip of the story. They glance at the pictures for 5 seconds and instantly understand the plot.
  • Why it helps: AI models can "read" an image much faster and cheaper than processing thousands of text tokens. It's a massive shortcut.

3. The "Ruler" (Agentless Layout Violation Detection)

When the AI generates the poster, it sometimes messes up. Text might run off the edge, or a box might be empty.

• The Old Way: Previous systems asked the AI to "look" at the poster and say, "Is this messy?" But the AI is bad at math and geometry, so it often missed the errors or got confused.
• The New Way: This module uses a simple, deterministic math algorithm (like a ruler and a protractor). It doesn't "think" or "guess"; it simply measures the pixels.
• The Analogy:
  • Old way: Asking a poet to measure a table with a ruler. They might say, "It feels about right," but they could be wrong.
  • New way: Using a laser level. It instantly tells you, "This line is 2 inches too long." It's fast, free, and 100% accurate.

The Result

By combining these three tricks, EfficientPosterGen achieves three amazing things:

1. It's Cheaper: It uses about 10 times less money (tokens) than previous methods.
2. It's Faster: Because it skips the boring parts and uses image shortcuts, it finishes the job much quicker.
3. It's Better: The posters it makes are clean, fit perfectly on the board, and actually look like professional academic posters rather than messy word documents.

In a nutshell:
If previous AI poster makers were like a clumsy intern trying to cram a library into a shoebox, EfficientPosterGen is like a professional architect who knows exactly which bricks to keep, how to pack them efficiently, and has a laser guide to ensure the wall is perfectly straight.

Technical Summary

1. Problem Statement

Automated academic poster generation aims to distill lengthy research papers into concise, visually coherent presentations. While Multimodal Large Language Models (MLLMs) have shown promise in this domain (e.g., PosterAgent), existing approaches suffer from three critical limitations:

1. Low Information Density & Redundancy: Feeding full academic papers (often 20k+ tokens) into MLLMs introduces substantial irrelevant content (references, acknowledgments, repetitive descriptions). This dilutes attention mechanisms, leading to posters that lack clear focal points or overemphasize secondary details.
2. Excessive Token Consumption: Processing full-text inputs in a pure textual modality approaches or exceeds the context window limits of many models (e.g., Llama3-8B is limited to 8k tokens). This incurs high computational costs, latency, and financial overhead, hindering scalability.
3. Unreliable Layout Verification: Current methods rely on auxiliary MLLMs to detect layout violations (e.g., text overflow, spatial sparsity). However, MLLMs often struggle with precise spatial reasoning and element localization, leading to unreliable detection and unnecessary regeneration cycles, further increasing token costs.

2. Methodology: EfficientPosterGen

The authors propose EfficientPosterGen, an end-to-end framework designed to reduce token costs while maintaining high poster quality. It consists of three core modules:

A. Semantic-aware Key Information Retrieval (SKIR)

To address low information density, SKIR filters redundant content before generation.

• Paragraph Grouping: Uses perplexity-based analysis to identify semantic boundaries within the text, grouping paragraphs into coherent segments.
• Semantic Graph Construction: Constructs a directed graph where nodes represent content segments. Edges are weighted by perplexity reduction, quantifying how much one segment contributes to the predictability (and thus semantic understanding) of another.
• Diversity-aware Selection: Employs a modified PageRank algorithm on the reversed graph to calculate semantic importance scores. To ensure diverse coverage across the paper's structure, a penalty factor based on the Lowest Common Ancestor (LCA) in the document's section tree is applied. This prevents the selection of segments clustered in a single section, ensuring a balanced representation of the paper.

B. Visual-based Context Compression (VCC)

To mitigate token limits and costs, VCC shifts the input modality from text to images.

• Text-to-Image Rendering: Selected high-value text segments are rendered into PNG images.
• Token Efficiency: By feeding these images to the MLLM instead of raw text, the framework significantly reduces token usage (studies show ~50% reduction) while preserving semantic readability. The MLLM then generates concise, poster-ready bullet points and configuration parameters (e.g., font size) based on these visual inputs.

C. Agentless Layout Violation Detection (ALVD)

To replace unreliable MLLM-based verification, ALVD uses a deterministic, algorithmic approach.

• Gradient Analysis: The generated panel images are divided into horizontal and vertical strips. The algorithm computes color gradient magnitudes along these strips.
• Content Region Detection: Strips with high gradient magnitudes (indicating text or image content) are activated. The Cartesian product of activated strips defines the content regions.
• Violation Logic:
  • Overflow: Detected if the minimum enclosing bounding box of the content exceeds the panel boundaries.
  • Sparsity: Detected if the ratio of the content area to the panel area falls below a threshold.
  • Validity: If neither condition is met, the layout is valid.
• Feedback Loop: If a violation is detected, the system triggers a regeneration of the specific panel with adjusted parameters (e.g., reducing font size for overflow) without invoking an MLLM for the check itself.

3. Key Contributions

1. EfficientPosterGen Framework: An end-to-end system that significantly reduces generation costs (token usage) while maintaining high-quality academic posters.
2. SKIR Module: A novel graph-based content extraction strategy that models semantic contributions and structural diversity to identify salient information, ensuring the poster covers the paper's core contributions without redundancy.
3. VCC Module: A visual compression technique that replaces long-context text inputs with image-based representations, effectively bypassing token limits and reducing computational overhead.
4. ALVD Algorithm: A deterministic, agentless layout verification method that outperforms MLLM-based detection in accuracy and speed, eliminating the token overhead associated with iterative layout checks.

4. Experimental Results

The framework was evaluated against baselines including PosterAgent (using GPT-5 and Qwen3 backbones) and end-to-end HTML generation methods.

• Token Efficiency: EfficientPosterGen achieved a ~10x reduction in token consumption compared to PosterAgent.
  • Ours-5 (GPT-5 backbone): ~21.38K tokens vs. PosterAgent's ~254K tokens.
  • Ours-Qwen (Qwen3 backbone): ~10.33K tokens vs. PosterAgent's ~125K tokens.
• Layout Reliability: The deterministic ALVD algorithm achieved 94% accuracy in detecting overflow and sparsity, significantly outperforming MLLM-based baselines (which ranged from 62% to 72%). It also reduced detection latency by ~33–49x.
• Poster Quality:
  • Visual Similarity: Achieved the highest scores (77.98% for Ours-5) compared to ground truth posters.
  • PaperQuiz (Content Comprehension): Outperformed baselines in both verbatim and interpretive question answering, with Ours-5 scoring 119.51 vs. PosterAgent-5's 110.02.
  • Human Evaluation: In a preference study, EfficientPosterGen was preferred over PosterAgent in 66.7% (GPT-5) and 73.3% (Qwen) of cases.
• Ablation Studies: Removing any of the three modules (SKIR, VCC, or ALVD) resulted in degraded performance. Notably, removing ALVD caused token consumption to surge to ~85K tokens due to inefficient MLLM-based regeneration loops.

5. Significance

EfficientPosterGen addresses the scalability bottleneck in automated academic poster generation. By decoupling semantic extraction from generation and replacing probabilistic MLLM verification with deterministic algorithms, the framework offers a cost-effective, reliable, and scalable solution. It demonstrates that high-quality multimodal generation does not require massive token inputs or iterative MLLM feedback loops, paving the way for industrial-scale deployment of automated academic communication tools. The code is open-source, facilitating further research in efficient multimodal generation.