Glyph-ByT5-v2: A Strong Aesthetic Baseline for Accurate Multilingual Visual Text Rendering This paper presents Glyph-ByT5-v2 and Glyph-SDXL-v2, which achieve accurate multilingual visual text rendering for 10 languages. The key contributions include creating a large multilingual glyph-text and graphic design dataset, building a multilingual visual paragraph benchmark, and leveraging step-aware preference learning to enhance visual aesthetics. The approach involves translating English glyph images and graphic design images into those for other languages, and using a multilingual text encoder and graphic generation model to render text in multiple languages. The models are trained on a large dataset of glyph-text pairs and graphic design images, and the visual aesthetics are improved using advanced techniques like SPO-SDXL. The results show that Glyph-SDXL-v2 outperforms previous models and commercial products like DALL·E3 in terms of visual text quality and aesthetics. The paper also includes a user study comparing the generated images with DALL·E3, showing that Glyph-SDXL-v2 is preferred by users. The approach provides a strong aesthetic baseline for accurate multilingual visual text rendering and inspires further research in this area.Glyph-ByT5-v2: A Strong Aesthetic Baseline for Accurate Multilingual Visual Text Rendering This paper presents Glyph-ByT5-v2 and Glyph-SDXL-v2, which achieve accurate multilingual visual text rendering for 10 languages. The key contributions include creating a large multilingual glyph-text and graphic design dataset, building a multilingual visual paragraph benchmark, and leveraging step-aware preference learning to enhance visual aesthetics. The approach involves translating English glyph images and graphic design images into those for other languages, and using a multilingual text encoder and graphic generation model to render text in multiple languages. The models are trained on a large dataset of glyph-text pairs and graphic design images, and the visual aesthetics are improved using advanced techniques like SPO-SDXL. The results show that Glyph-SDXL-v2 outperforms previous models and commercial products like DALL·E3 in terms of visual text quality and aesthetics. The paper also includes a user study comparing the generated images with DALL·E3, showing that Glyph-SDXL-v2 is preferred by users. The approach provides a strong aesthetic baseline for accurate multilingual visual text rendering and inspires further research in this area.