DesignLab fine-tunes large language models for the design reviewer and contributor roles and simulates intermediate drafts by introducing controlled perturbations. This enables the design reviewer to learn design errors and the contributor to learn how to fix them. The system's iterative process continues until the design is polished to perfection. The design reviewer identifies design issues using the TENTATIVE label, and the design contributor suggests improvements, which are repeated until the design has no more issues.
DesignLab's effectiveness is demonstrated through a user study that evaluates the effect of its iterative refinement on slide aesthetics. The results show that DesignLab significantly improves the aesthetics of presentation slides in each iteration, outperforming a commercial tool and a state-of-the-art design generation method. DesignLab's ability to polish rough ideas into professional-looking presentations makes it an valuable tool for individuals who struggle with creating high-quality presentation slides.
