The Diffusion CNN (DiC) is a scalable pure 3×3 convolution diffusion model proposed by Huawei, Peking University, and Beijing University of Posts and Telecommunications on June 8, 2025. It is a diffusion model architecture baseline that combines speed and performance. The related paper results are "DiC: Rethinking Conv3x3 Designs in Diffusion Models".

Experiments at various scales and settings show that DiC significantly outperforms the existing diffusion model Transformer while maintaining a good inference speed advantage.