Abstract
Creative graphic design involves iterative refinement over structured elements, where subtle spatial and typographic defects can harm readability and visual quality. Although recent vision-language models can generate plausible layouts, reliable fine-grained correction remains difficult due to constraint violations and the ambiguity of valid edits. We introduce DesignCorrection-R1, a reasoning-oriented vision-language policy that performs local, intent-preserving repair by generating minimal executable edit programs over a structured element-level action space. The model is trained with a staged curriculum for layout and style repair and optimized using DeepSeek-R1-inspired reinforcement learning with constraint-based and minimal-edit rewards. Preliminary results on synthetic layout defects show improved issue localization, supporting the effectiveness of structured reasoning for design correction.
Shunsuke Kitada, Ph.D.
Research Scientist working on Vision & Language with Deep Learning
My research interests include deep learning-based natural language processing, computer vision, medical image processing, and computational advertising.