A rede residual é vista como uma denoising discreto, a correspondência de pontuação resolve o treino em blocos, sem precisar de comunicação entre blocos, e o uso de memória é muito menor.

DiffusionBlocks divide Transformer-based networks into independently trainable blocks, treating residual networks as discretized denoising steps, utilizing score matching to achieve block-level training without inter-block communication, significantly reducing training memory. Experiments show effectiveness across multiple architectures, with only one block activated per step during inference, reducing the computational cost of 12-layer DiT (B=3) to one-third of the original. This method is suitable for ViT/DiT/MDM/AR Transformer, but requires input-output dimension matching and cannot be used with U-Net.