Research map / Image generation / Text-to-image diffusion
TextAtlas5M: A Large-Scale Dataset for Long Text Image Generation
Dongxing Mao, Alex Jinpeng Wang, Weiming Han, Jiawei Zhang, Zhuobai Dong, Linjie Li, Yiqi Lin, Zhengyuan Yang, Libo Qin, Fuwei Zhang, Lijuan Wang, Min Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Lumina-mGPT 7B (1024) on TextAtlas5M | Lumina-mGPT 7B · TextAtlasEval CleanTextSynth | OCR accuracy: 0.32 -> 48.55 | Improved |
| fine-tune Lumina-mGPT 7B (1024) on TextAtlas5M | Lumina-mGPT 7B · TextAtlasEval TextVisionBlend | OCR accuracy: 0.03 -> 65.34 | Improved |
| fine-tune PixArt-alpha 0.6B (512) on TextAtlas5M | PixArt-alpha 0.6B · TextAtlasEval CleanTextSynth | OCR accuracy: 0.23 -> 39.14 | Improved |
| fine-tune Janus-Pro 1B on TextAtlas5M | Janus-Pro 1B · TextAtlasEval CleanTextSynth | OCR accuracy: 0.27 -> 5.23 | Improved |
| fine-tune on TextAtlas5M, hard real-image subset | Lumina-mGPT 7B (1024) · TextAtlasEval TextScenesHQ | OCR accuracy: 0.22 -> 6.44 | Improved |
| train on TextAtlas5M instead of Marion10M (ablation) | PixArt-alpha 0.6B (512) · LAIONEval4000 (short text) | OCR accuracy: 52.37 -> 67.53 | Improved |
| add synthetic subsets to real-only training data (ablation) | Lumina-mGPT 7B · TextScenesHQ Eval | OCR accuracy: 5.32 -> 6.44 | Improved |
| raise Lumina-mGPT resolution from 512 to 1024 after fine-tuning (ablation) | Lumina-mGPT 7B fine-tuned · TextAtlasEval CleanTextSynth | OCR accuracy: 32.90 -> 48.55 | Improved |
| fine-tune Lumina-mGPT 7B (512) on TextAtlas5M | Lumina-mGPT 7B (512) · TextAtlasEval TextVisionBlend CLIP score | CLIP score: 0.19 -> 0.16 | Hurt |
| fine-tune on TextAtlas5M, StyledTextSynth subset | Lumina-mGPT 7B (512) · TextAtlasEval StyledTextSynth | OCR accuracy: 0.09 -> 5.43 | Improved |
| use lower-resolution discrete VQ tokenizer for AR text rendering (ablation) | Janus-Pro VQ-VAE tokenizer · image reconstruction | text fidelity | Hurt |
| increase input resolution of VQ-VAE tokenizer (256 to 1024) (ablation) | Janus-Pro VQ-VAE tokenizer · image reconstruction | text clarity | Improved |
Closest papers by meaning
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.89)
- Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis (CVPR 2025 · similarity 0.89)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.88)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.88)
- Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.88)
- OpenING: A Comprehensive Benchmark for Judging Open-ended Interleaved Image-Text Generation (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
