Redes residuais vistas como denoising discreto, correspondência de pontuação resolve o treinamento em blocos, sem precisar de comunicação entre blocos, podendo treinar, com muito menos pressão de memória.

DiffusionBlocks divide the Transformer-based network 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 and output dimensions to match and cannot be used with U-Net.