PosterCraft – An Aesthetic Poster Generation Framework Jointly Developed by HKUST and Meituan

What is PosterCraft？

PosterCraft is a unified framework for generating high-quality, aesthetically pleasing posters, developed by The Hong Kong University of Science and Technology (Guangzhou) in collaboration with Meituan and other institutions. Unlike traditional modular workflows and fixed template layouts, PosterCraft allows models to freely explore coherent and visually engaging compositions. The framework features a four-stage cascaded workflow that optimizes poster generation through scalable text rendering, high-quality fine-tuning, aesthetic text-based reinforcement learning, and vision-language refinement. Each stage is supported by an automated data pipeline tailored to its specific goal, enabling robust training without requiring complex architectural modifications. In extensive experiments, PosterCraft significantly outperforms open-source baselines in text rendering accuracy, layout coherence, and overall visual appeal—approaching the quality of commercial systems.

Key Features of PosterCraft

• High-Quality Text Rendering: Accurately renders text, ensuring clarity and correctness of content.

• Artistic Content Generation: Produces visually rich and abstract artistic elements, giving posters a unique aesthetic style.

• Striking Layout Design: Creates visually impactful and coherent layouts that enhance overall design consistency.

• End-to-End Generation: From text input to final poster creation, the entire process is completed within a single model, without relying on external modules or preset templates.

• Aesthetic Optimization: Utilizes reinforcement learning and vision-language feedback mechanisms to enhance both the aesthetic quality and content accuracy of the posters.

Technical Principles Behind PosterCraft

• Scalable Text Rendering Optimization: Trained on the large-scale Text-Render-2M dataset to improve the clarity and accuracy of text rendering.

• High-Quality Poster Fine-Tuning: Supervised fine-tuning with the HQ-Poster-100K dataset to enhance visual quality and maintain a consistent artistic style.

• Aesthetic Text-Based Reinforcement Learning: Trained with the Poster-Preference-100K dataset to guide the model towards outputs that align better with human aesthetic preferences.

• Vision-Language Feedback Refinement: Uses the Poster-Reflect-120K dataset and multimodal feedback to further refine generated posters, improving both content precision and aesthetic appeal.

PosterCraft Project Links

• Official Website: https://ephemeral182.github.io/PosterCraft/

• GitHub Repository: https://github.com/Ephemeral182/PosterCraft

• Hugging Face Model Hub: https://huggingface.co/PosterCraft

• arXiv Technical Paper: https://arxiv.org/pdf/2506.10741

Application Scenarios for PosterCraft

• Movie Posters: Generate eye-catching posters based on film themes, highlighting key elements and visual impact.

• Art Exhibition Posters: Create culturally rich and artistically styled posters that reflect the tone and message of the exhibition.

• Product Promotion Posters: Generate promotional posters that emphasize product features and advantages in a visually appealing way.

• Academic Conference Posters: Design professional posters with a scholarly tone that effectively communicate conference themes and agendas.

• Campus Event Posters: Produce creative and engaging posters to promote student and university events, showcasing content highlights.