LeX-Art: Rethinking Text Generation via Scalable High-Quality Data Synthesis
LeX-Art enhances text rendering in T2I models via high-quality data synthesis, achieving 79.81% PNED improvement and superior attribute accuracy.
Key Findings
Methodology
LeX-Art employs a data-centric approach, leveraging Deepseek-R1 for prompt refinement, constructing LeX-10K high-quality dataset through multi-stage filtering and knowledge-augmented recaptioning. Fine-tuning LeX-FLUX and LeX-Lumina models on this dataset, the framework introduces PNED for robust text accuracy evaluation. The system integrates prompt enhancement, dataset curation, and model optimization, resulting in significant improvements in text fidelity and aesthetics, validated through comprehensive benchmarks.
Key Results
- LeX-Lumina achieved a 79.81% increase in PNED on CreateBench, demonstrating substantial accuracy gains in text rendering.
- LeX-FLUX outperformed baselines in color (+3.18%), position (+4.45%), and font accuracy (+3.81%), confirming enhanced attribute control.
- The curated LeX-10K dataset, combined with multi-stage filtering, led to more aesthetically refined and layout-diverse images, outperforming existing datasets in quality scores.
Significance
This work addresses longstanding challenges in visual text generation by shifting from control-based to data-driven strategies, enabling models to produce more natural, accurate, and aesthetically pleasing text images. It bridges the gap between prompt expressiveness and rendering fidelity, impacting industries like digital advertising, branding, and artistic design. The introduced evaluation metrics and benchmark set new standards for assessing visual text quality, fostering further research in scalable high-quality synthesis.
Technical Contribution
The core innovation lies in integrating Deepseek-R1 for prompt enhancement, multi-stage filtering for dataset quality, and the development of PNED for flexible accuracy measurement. Fine-tuning large models like FLUX.1 and Lumina-Image 2.0 on LeX-10K, the framework achieves state-of-the-art text rendering. The approach emphasizes data quality over architectural modifications, opening new avenues for scalable, high-fidelity text synthesis.
Novelty
This is the first comprehensive system combining deep prompt enhancement, rigorous dataset construction, and a novel evaluation metric (PNED) to significantly improve non-glyph-conditioned text rendering. Unlike prior works relying heavily on glyph modules, this approach leverages data quality and model fine-tuning, setting a new paradigm for scalable, aesthetic visual text generation.
Limitations
- The models still struggle with extremely complex layouts or rare fonts, mainly due to limitations in training data diversity and computational resources.
- High-quality dataset creation involves substantial manual effort and computational cost, hindering rapid scaling.
- Generalization across multiple languages and cultural contexts remains limited; future work needs to incorporate multilingual, multi-font data for broader applicability.
Future Work
Future directions include automating dataset扩展，结合多模态信息提升多语言、多场景适应性，引入强化学习优化布局和风格控制，增强模型的交互性和个性化能力。还将探索多任务学习和生成对抗网络（GAN）结合，进一步提升文本与图像的自然融合和美学表现。
AI Executive Summary
LeX-Art introduces a scalable framework for high-quality visual text generation, addressing the persistent challenge of accurate and aesthetic text rendering in T2I models. By leveraging Deepseek-R1 for prompt refinement, the authors construct LeX-10K—a dataset of 10,000 high-resolution, aesthetically refined image-caption pairs through multi-stage filtering and knowledge-augmented recaptioning. This dataset forms the foundation for fine-tuning two models, LeX-FLUX and LeX-Lumina, which demonstrate state-of-the-art performance in text attribute accuracy and fidelity.
The core innovation lies in shifting from architectural control modules to a data-centric paradigm, emphasizing the importance of high-quality training data. The introduction of PNED as a flexible, robust metric for text accuracy evaluation further advances the field, enabling comprehensive assessment beyond traditional OCR-based measures. Experimental results show that LeX-Lumina improves PNED by 79.81% on CreateBench, while LeX-FLUX surpasses baselines in color, position, and font accuracy.
This approach significantly enhances the ability of models to generate complex, multi-word text images with diverse layouts and styles, broadening their application scope in design, advertising, and art. The framework’s scalability and robustness open new avenues for future research, including multilingual and multi-font synthesis, multi-modal integration, and user-interactive generation systems. Despite these advances, challenges remain in further reducing computational costs and expanding dataset diversity. Overall, LeX-Art marks a substantial step forward in scalable, high-fidelity visual text synthesis, promising impactful contributions to both academia and industry.
Deep Analysis
Background
随着深度学习模型在图像生成领域的突破，文本到图像（T2I）技术已成为研究热点。代表性工作如Stable Diffusion、DALL·E推动了高质量图像生成，但在文本渲染方面仍存在不足。现有数据集如AnyWord-3M和MARIO-10M在分辨率和多样性方面有限，导致模型在细节控制和美学表现上受限。近年来，控制模块如GlyphControl、TextDiffuser试图引入glyph信息改善文本效果，但在自然融合和多样性方面仍不足。整体背景显示，提升文本渲染质量的关键在于高质量数据和模型微调的结合，亟需系统性解决方案。
Core Problem
核心问题在于现有模型难以在复杂布局、多词、多样字体场景中实现高精度文本渲染，受限于数据质量和模型泛化能力。传统控制方法虽能保证一定的文本准确性，但在美学和布局灵活性方面表现不足，尤其在多语言、多场景环境中。如何在保证文本内容正确的基础上，提升整体视觉效果和布局自然性，成为亟待攻克的难题。
Innovation
本研究提出三大创新：第一，基于Deepseek-R1的提示增强机制，丰富文本描述细节；第二，构建LeX-10K高质量数据集，采用多阶段筛选和知识重述确保多样性和准确性；第三，引入PNED指标，增强文本准确性评估的鲁棒性。这些创新结合数据驱动和模型微调，突破了传统依赖控制模块的局限，显著提升文本渲染的自然度和细节表现。
Methodology
- �� 利用Deepseek-R1对原始提示进行细粒度增强，加入字体、颜色、布局信息。
- �� 通过多阶段筛选，包括Q-Align和Paddle-OCR，筛除低质量和不符合要求的图像。
- �� 采用知识增强的GPT-4o对文本进行重述，确保与图像内容高度一致。
- �� 构建LeX-10K数据集，结合筛选和重述，确保高质量、多样性。
- �� 微调LeX-FLUX（基于FLUX.1）和LeX-Lumina（基于Lumina-Image 2.0），提升文本渲染能力。
- �� 引入PNED指标，用于评估生成文本与输入提示的匹配度，增强评估的全面性。
Experiments
采用LeX-10K作为微调数据，在CreateBench、SimpleBench和AnyText-Benchmark上进行评估。指标包括OCR Recall、PNED、CLIP Score和FID。训练参数如LeX-FLUX的学习率1e-6，批次256，6000步。通过对比Glyph条件模型，验证数据增强的有效性。还进行用户偏好和VQA评估，确保多维性能。
Results
LeX-Lumina在CreateBench上PNED指标提升79.81%，显示出极强的文本准确性。LeX-FLUX在色彩、位置、字体准确率上分别超越基线3.18%、4.45%、3.81%。多项指标验证了数据增强和微调策略的有效性，模型在复杂布局、多词场景中表现优异，验证了高质量数据和模型微调的结合优势。
Applications
该技术适用于广告设计、品牌标识、艺术创作、动态海报等场景，满足对高质量、多样化文本图像的需求。未来结合多语种、多字体、多场景数据，将推动AI在数字媒体、虚拟现实等领域的深度融合。
Limitations & Outlook
模型在极端复杂布局或少样本场景下仍存在模糊或错位问题，主要受限于数据多样性和模型容量。高质量数据集的构建成本较高，难以快速扩展。未来需优化模型结构和多模态融合能力，提升泛化能力。
Plain Language Accessible to non-experts
想象你在一家厨房做菜，食材代表数据，厨师代表模型。过去，厨师只能用有限的食材做出普通菜肴，效果不够漂亮也不够美味。现在，厨师得到了一份特别的食谱（高质量数据集），配料更丰富，步骤更详细。厨师按照新食谱，经过多次试验和调整，做出色香味俱佳的菜肴（高质量文本图像）。这个过程就像用LeX-Art的方法，通过优化提示和筛选，确保每一道菜都色彩鲜明、布局合理、味道鲜美。最终，这家厨房能做出各种复杂、精美的菜肴，满足不同顾客的需求。
ELI14 Explained like you're 14
想象你在玩拼图游戏，目标是拼出一幅漂亮的画。以前，拼图的图片不够清楚，拼出来的画总是模糊或错位。现在，有了新方法，就像给拼图图片加了标签和说明，让每块拼图都能准确拼到正确位置。科学家们用一种叫LeX-Art的技术，先用特别聪明的算法把提示变得更详细，再筛选出最清晰、最漂亮的图片。接着，他们训练两个“拼图师”模型，一个专注于高质量拼图（LeX-FLUX），另一个则更快、更轻便（LeX-Lumina）。他们还发明了一个新标准（PNED），用来判断拼图是否拼得像原图一样。结果显示，这些新技术能让拼出来的画既漂亮又准确，比以前的拼图效果好多了。就像你用更聪明的拼图指南，拼出了最棒的画！
Abstract
We introduce LeX-Art, a comprehensive suite for high-quality text-image synthesis that systematically bridges the gap between prompt expressiveness and text rendering fidelity. Our approach follows a data-centric paradigm, constructing a high-quality data synthesis pipeline based on Deepseek-R1 to curate LeX-10K, a dataset of 10K high-resolution, aesthetically refined 1024$\times$1024 images. Beyond dataset construction, we develop LeX-Enhancer, a robust prompt enrichment model, and train two text-to-image models, LeX-FLUX and LeX-Lumina, achieving state-of-the-art text rendering performance. To systematically evaluate visual text generation, we introduce LeX-Bench, a benchmark that assesses fidelity, aesthetics, and alignment, complemented by Pairwise Normalized Edit Distance (PNED), a novel metric for robust text accuracy evaluation. Experiments demonstrate significant improvements, with LeX-Lumina achieving a 79.81% PNED gain on CreateBench, and LeX-FLUX outperforming baselines in color (+3.18%), positional (+4.45%), and font accuracy (+3.81%). Our codes, models, datasets, and demo are publicly available.