Diffusion models excel at generating high-dimensional data, but their training efficiency and representation quality are inferior to self-supervised learning methods. This paper reveals that the lack of high-quality, semantically rich representations during training is a key bottleneck. Systematic analysis identifies a crucial representation processing region (early layer) where semantic and structural pattern learning primarily occurs before the model performs generation. To address this, we propose Embedded Representation Warmup (ERW), a plug-and-play framework that initializes the early layer of a diffusion model with high-quality, pre-trained representations, acting as a warmup. This warmup reduces the burden of learning representations from scratch, thereby accelerating convergence and improving performance. The effectiveness of ERW relies on its precise integration into a specific neural network layer (the representation processing region), where the model primarily processes and transforms feature representations for subsequent generation. ERW not only accelerates training convergence but also improves representation quality, experimentally achieving a 40x training speedup compared to the existing state-of-the-art method, REPA.
