Representation Conditioning (RECON): Accelerating Diffusion Transformer Training

Forum topic · 小凯 · 2026-10-16

Summary

Researchers Saeed Razavi, Xingjian Leng, and Liang Zheng propose Representation Conditioning (RECON), a method to improve the training efficiency of diffusion transformers in computer vision. Existing representation-alignment techniques like REPA use a pretrained vision encoder to supervise an intermediate layer via an MLP projection head, importing encoder semantics into the diffusion backbone; the head is discarded after training. However, while early layers benefit from the alignment loss, later layers supervised only by the velocity prediction loss have no guaranteed access to these semantics. RECON addresses this by feeding pretrained vision encoder tokens directly into later layers during training. At inference time, since clean images are unavailable, the MLP projection head is reused as a surrogate for the vision encoder and its output tokens are projected into the later layers. The approach adds no extra parameters and dramatically accelerates convergence: RECON trains 63x faster than SiT and 25x faster than REPA. The paper is available as arXiv 2610.20608.

Paper Overview

Field: Computer Vision (CV)
Authors: Saeed Razavi, Xingjian Leng, Liang Zheng
Published: 2026-10-14
arXiv: 2610.20608

Summary

Representation-alignment methods effectively accelerate the training of diffusion transformers. Taking REPA as an example, it uses a pretrained vision encoder to supervise an intermediate layer through an MLP projection head, which is discarded after training. The goal is to import semantics from the vision encoder into the diffusion backbone.

While this objective is achieved at early layers by back-propagating the alignment loss, there is no guarantee that later layers — supervised by the velocity prediction loss alone — can well access such semantics.

The authors propose a simple solution, representation conditioning (RECON), to improve semantics in later layers:

During training: the tokens from the pretrained vision encoder are fed into the later layers.
During inference: because clean images are no longer available, the MLP projection head is reused as a surrogate for the vision encoder, and its output tokens are projected into the later layers.

Key results

The method adds no extra parameters.
Convergence is much faster: RECON is 63x faster than SiT and 25x faster than REPA.

Original Abstract (excerpt)

> Representation-alignment methods effectively accelerate the diffusion transformer training. REPA, for example, uses a pretrained vision encoder to supervise an intermediate layer through an MLP projection head which is discarded after training. It aims to import semantics from the vision encoder to the diffusion backbone. While this objective is achieved at early layers by back-propagating the alignment loss, there is no guarantee that later layers, supervised by the velocity prediction loss alone, can well access such semantics. We propose a simple solution, representation conditioning (RECON), to improve semantics in later layers. During training, we feed the pretrained vision encoder tokens into later layers; during inference, because we no longer have the clean image, we reuse the MLP ...

*Auto-collected on 2026-10-16*

Representation Conditioning (RECON): Accelerating Diffusion Transformer Training

Summary

Paper Overview

Summary

Key results

Original Abstract (excerpt)

Tags