TextAtlas5M: A Large-scale Dataset for Dense Text Image Generation

Abstract: Text-conditioned image generation has gained significant attention in recent years and are processing increasingly longer and comprehensive text prompt. In everyday life, dense and intricate text appears in contexts like advertisements, infographics, and signage, where the integration of both text and visuals is essential for conveying complex information. However, despite these advances, the generation of images containing long-form text remains a persistent challenge, largely due to the limitations of existing datasets, which often focus on shorter and simpler text. To address this gap, we introduce TextAtlas5M, a novel dataset specifically designed to evaluate long-text rendering in text-conditioned image generation. Our dataset consists of 5 million long-text generated and collected images across diverse data types, enabling comprehensive evaluation of large-scale generative models on long-text image generation. We further curate 3000 human-improved test set TextAtlasEval across 3 data domains, establishing one of the most extensive benchmarks for text-conditioned generation. Evaluations suggest that the TextAtlasEval benchmarks present significant challenges even for the most advanced proprietary models (e.g. GPT4o with DallE-3), while their open-source counterparts show an even larger performance gap. These evidences position TextAtlas5M as a valuable dataset for training and evaluating future-generation text-conditioned image generation models.

• The paper introduces TextAtlas5M, a large dataset of 5 million images designed to train and evaluate models generating images with dense text content.
• TextAtlas5M includes synthetic images created in stages and real-world images like slides and book covers to represent diverse dense text scenarios.
• Evaluation shows current models struggle with dense text generation, indicating TextAtlas5M will be crucial for improving accuracy and fidelity in future models.

Overview of TextAtlas5M: A Large-scale Dataset for Dense Text Image Generation

The academic article explores the intricacies of text-conditioned image generation, with a particular focus on handling images composed of dense textual content. The paper introduces TextAtlas5M, a comprehensive dataset designed to address the shortcomings of existing datasets, which predominantly concentrate on simple, short text. The creation of TextAtlas5M aims to facilitate the generation and evaluation of images embedded with lengthy, intricate text, thereby setting a new benchmark for evaluating the capabilities of text-conditioned generative models.

Text-conditioned image generation has evolved to process increasingly lengthy text prompts, advancing from an input length of 77 text tokens in systems like Dall-E to a capacity of 2000 tokens in more recent autoregressive architectures. This progression paves the way for more comprehensive and dynamically controllable image generation based on substantial text input. However, the field still grapples with generating images containing extensive text due to the constraints imposed by current datasets.

Dataset Construction

TextAtlas5M comprises five million images generated or collected to encompass diverse data types, such as synthetic and real-world datasets. The dataset has been assembled with two pivotal components: synthetic images and real images.

1. Synthetic Images: These are crafted in three stages—beginning with simple text on plain backgrounds and advancing to synthetic representations of interleaved data, culminating in the integration of text within complex synthetic-natural images.
2. Real Images: This subset incorporates images from real-world scenarios, including slides from academic presentations, book covers, and web-sourced dense-text imagery. The acquisition process integrates manual curation to ensure high-quality data.

The dataset incorporates a refined evaluation benchmark, TextAtlasEval, which spans multiple domains, offering a platform to assess the generation capabilities of both proprietary and open-source models. The varying levels of complexity in TextAtlas5M make it a valuable resource for training future models and fortifying their ability to regenerate dense text images accurately.

Evaluation and Implications

The study assesses several contemporary text-to-image generation models using the TextAtlasEval benchmark. Notably, the study highlights stark performances of models such as GPT4o with DallE-3 compared to open-source counterparts, identifying substantial inaccuracies and limitations in rendering long-form text.

TextAtlas5M sets a precedent not just in evaluating the functional accuracy of current generation models, but also in specifying a path toward future innovations. The dataset's multifaceted nature—rooted in diverse real-world embeddings—positions it as pivotal for the development of algorithms capable of rendering comprehensive text in images with high fidelity and cogency.

With its public availability, TextAtlas5M offers an expansive resource that could cultivate advancements both in research and practical applications. It is anticipated that future endeavors in AI modeling will leverage such a dataset to improve models' understanding and synthesis capabilities, ensuring accurate text replication across dynamic, visually stimulating layouts.

Conclusion

By introducing such a robust dataset, the authors provide a pertinent contribution to the domain of AI and image generation, facilitating growth in both technical robustness and application-specific accuracy. The insights garnered from the evaluations underscore ongoing challenges while illuminating potential routes for refined models that can deftly handle dense text image synthesis. The extensiveness and applicability of TextAtlas5M make it a cornerstone resource, capable of influencing both theoretical research and real-world applications in text-to-image generation.