Yaxin Liu¹ (), Maxwell Kay¹, Kibum Moon¹, Roger Beaty², Adam Green¹; ¹Georgetown University, ²The Pennsylvania State University

Recent advances in generative artificial intelligence (AI) have positioned it as a black-box parallel to the human mind, enabling comparisons between artificial and human cognition. However, much remains unknown about how the creativity of these models compares to human creativity, a hallmark of intelligence. In particular, the ability to generate original outputs that transcend initial perceptual constraints or patterns (i.e., overcoming fixations) remains largely understudied. The present study assessed creative variability by comparing humans and image-generation models using the Multi-Trial Creative Ideation (MTCI) "incomplete shapes" task. In this task, participants generated original doodles by incorporating existing incomplete shapes and lines. Drawings from human participants were compared with outputs from two types of image-generation models: diffusion-based (i.e., Stable Diffusion Img2Img pipeline) and vision transformer-based (i.e., Dalle-E). Both models were configured to simulate line-drawing outputs and were uniformly prompted to generate doodles using the provided incomplete shapes. We analyzed the visual creativity and perceptual similarity of over 6000 drawings. Specifically, we computationally assessed the visual creativity of each drawing by using the Automated Drawing Assessment (AuDra) model, a convolutional neural network (CNN) trained on and validated against human ratings (Patterson et al., 2023). We also quantified the perceptual similarity of drawings using metrics such as the structural similarity index (SSIM) and perceptual hash. Pairwise perceptual similarity scores were averaged across image data. Our findings revealed that human drawings consistently scored higher in visual creativity, as assessed by AuDra, compared to AI drawings. Furthermore, AI drawings exhibited greater perceptual similarity, indicating higher homogeneity compared to those generated by humans. Interestingly, image-generation models were sensitive to changes in hyperparameters. These results provide insight into how humans and generative AI leverage visual inputs to produce creative outputs and may reflect distinct processes in assimilating and generating visual information.