What is PosterCraft
PosterCraft is a unified framework launched by the Hong Kong University of Science and Technology (Guangzhou) and Meituan and other institutions for generating high-quality aesthetic posters. The framework abandons modular design processes and fixed predefined layouts, and supports the model to freely explore coherent and visually engaging compositions. The framework optimizes high-aesthetic poster generation based on a cascaded workflow with four key stages: scalable text rendering optimization, high-quality poster fine-tuning, reinforcement learning based on aesthetic text, and vision-language feedback refinement. Each stage is supported by a dedicated automated data construction pipeline to meet specific needs and enable robust training without complex architectural modifications. In multiple experiments, PosterCraft significantly outperforms open-source baselines in rendering accuracy, layout coherence, and overall visual appeal, approaching the quality of commercial systems.
Main Features of PosterCraft
- High-quality text rendering: Can accurately render text, ensuring the clarity and accuracy of text content.
- Artistic content creation: Generates visual content with an abstract artistic feel, giving posters a unique aesthetic style.
- Eye-catching layout design: Creates layouts with visual impact, ensuring the coordination and consistency of the overall design.
- End-to-end generation: From text input to the generation of the final poster, the entire process is completed in a single model, without external modules or preset templates.
- Aesthetic optimization: Based on reinforcement learning and vision-language feedback mechanisms, optimizes the aesthetic quality and content accuracy of posters.
Technical Principles of PosterCraft
- Scalable text rendering optimization: Based on the large-scale Text-Render-2M dataset, performs optimization training on the model for text rendering to improve the accuracy and clarity of text.
- High-quality poster fine-tuning: Performs supervised fine-tuning based on the HQ-Poster-100K dataset to improve the overall visual quality of posters and the consistency of artistic style.
- Reinforcement learning based on aesthetic text: Uses the Poster-Preference-100K dataset and, based on best preference optimization, trains the model to generate posters that better match human aesthetics.
- Vision-language feedback refinement: With the help of the Poster-Reflect-120K dataset, combines multimodal feedback to further optimize and adjust the generated posters, improving content accuracy and aesthetic value.
Project Address of PosterCraft
- Project official website：https://ephemeral182.github.io/PosterCraft/
- GitHub repository：https://github.com/Ephemeral182/PosterCraft
- HuggingFace model library：https://huggingface.co/PosterCraft
- arXiv technical paper：https://arxiv.org/pdf/2506.10741
Application Scenarios of PosterCraft
- Movie posters: Generates attractive posters according to the movie theme, highlighting key elements and visual impact.
- Art exhibition posters: Generates posters with an artistic feel and cultural atmosphere, showcasing the exhibition concept and style characteristics.
- Product promotion posters: Generates attractive promotional posters according to product features, showcasing functions and advantages.
- Academic conference posters: Generates posters with a professional and academic atmosphere, showcasing the conference theme and agenda.
- Campus activity posters: Produces creative posters, showcasing activity content and highlights.