This paper broadens the scope of score-based generative models (SGMs) through mean-field theory to improve the handling of high-cardinality data distribution. Specifically, we introduce mean-field chaotic diffusion models (MF-CDMs) that leverage propagation of chaos property for efficient management of infinitely many complex particle systems, aiming to address issues of the curse of dimensionality. We propose a novel score-matching framework in the infinite-dimensional chaotic particle system and the approximation scheme with a subdivision strategy for ease of computational complexity and efficient training. Our theoretical results verify asymptotically robust performance over growing complexity based on mean-field analysis, a property not achieved by existing methods. Empirical results confirm the superiority of our approach, showing consistent robustness on large cardinality data.