Representation-alignment methods like REPA accelerate diffusion transformer training by using a pretrained vision encoder to supervise an intermediate layer through an MLP projection head. That alignment loss is back-propagated to early layers, but later layers, trained only with the velocity prediction loss, may not fully access the imported semantics. The new paper proposes representation conditioning (RECON) to address this gap.
During training, RECON feeds the pretrained vision encoder's tokens directly into later layers. At inference, the clean image is no longer available, so the method reuses the MLP head as a proxy for the encoder and projects its output tokens into those same later layers. This adds no parameters and, according to the authors, leads to much faster convergence: 63× faster than SiT and 25× faster than REPA.
The paper is a single source, so there are no conflicting findings to compare. The authors provide code and a project page, and the work is posted on arXiv under the Computer Vision and Pattern Recognition category.