Building the denoising network of a diffusion model from 3x3 convolutions instead of self-attention transformers, to keep the scalability/quality of modern diffusion while avoiding the quadratic cost and slow inference of attention. Covers why a U-Net-like encoder-decoder hourglass outperforms an isotropic conv stack, how sparse skip connections cut redundancy and restore scaling, and how to inject class/timestep conditioning into a conv net (stage-specific embeddings, mid-block condition injection, conditional gating). Exemplified by DiC, which beats diffusion transformers (DiT) at comparable compute with a speed advantage.
DiC (Diffusion CNN) is a purely convolutional diffusion-model backbone that challenges the trend toward fully transformer-based (isotropic) diffusion architectures like DiT, whose self-attention give…