@Ryrenz: An open-source model specifically for poster generation, developed by HKUST and Meituan. Corresponding paper arXiv 2506.10741, with 14 authors from HKUST Guangzhou, Meituan, Xiamen University, and NUS. AI image generation has advanced rapidly in recent years, but poster generation has always been a problem area: the image can be created, but adding Chinese titles often reveals flaws—missing strokes, misaligned text…

An open-source model specifically for poster generation, developed by HKUST and Meituan. Corresponding paper arXiv 2506.10741, with 14 authors from HKUST Guangzhou, Meituan, Xiamen University, and NUS. AI image generation has made rapid progress in recent years, but poster generation has been a major challenge: the image can be generated, but adding Chinese titles often leads to issues like missing strokes and misaligned text, forcing users to go back to Photoshop to re-edit. PosterCraft addresses this problem, with training divided into four stages: first, text rendering optimization; second, fine-tuning with high-quality posters; third, aesthetic text reinforcement learning; and fourth, visual-language feedback. It includes the release of two models, PosterCraft-v1_RL and v1_Reflect, along with four datasets: Text-Render-2M, HQ-Poster-100K, Poster-Preference-100K, and Poster-Reflect-120K. Inference code and a Gradio interface are provided, along with a CPU offloading version for machines with limited GPU memory. It's worth noting honestly that on the benchmarks listed by the project, the text recall, F-score, and accuracy are 0.787, 0.774, and 0.735 respectively, while Gemini2.0-Flash-Gen has slightly higher scores for all three in the same table. Its value lies in the open weights and complete datasets, not in being the absolute top performer. Models that can be deployed and fine-tuned by users often have greater long-term value. GitHub: