cs.CVNov 18, 2026

Where Should a Diffusion Transformer Meet Its Teacher?

Authors: Zhuoqi Hu, Shiwei Li, Haozhao Wang, Xiandi Luo, Yihao Ouyang, Wenting Shen, Ang Wang, Jiamang Wang, 

Organizations: Huazhong University of Science and Technology · Alibaba Group · Shanghai Jiao Tong University

Abstract

Representation Alignment (REPA) significantly accelerates Diffusion Transformer training by guiding intermediate student features with a pretrained visual teacher, yet existing methods typically attach supervision to a single fixed layer. As student representations evolve, however, we find that the most effective layer for alignment varies across training stages. To track this evolving layer suitability online without costly search, we measure gradient compatibility between the flow-matching and alignment objectives, and propose AdaREPA, an online routing framework that adaptively selects the supervision layer as training progresses. AdaREPA uses a softmax distribution over smoothed layer compatibility scores to sample a layer for alignment at each training step. This probabilistic routing adapts teacher supervision to the student's changing learning needs, with minimal training overhead and no additional inference cost. On ImageNet-256, AdaREPA accelerates convergence and improves generation quality across model sizes, while revealing how the most suitable layers for supervision shift over training.

Figures & tables

Explore similar work

Sep 28, 2026

What Visual Generators Need from Teachers: Rethinking Representation Alignment

Representation alignment speeds up diffusion transformer training by pulling an intermediate block of the model (student) toward features of a frozen pretrained encoder (teacher). Which teacher layer to align, and for how long, is still set by convention, and each alternative costs a training run. We find that alignment helps where the student cannot linearly recover the teacher's features, not where it already resembles them. Since a deep teacher layer is largely predictable from the one below, we isolate what each layer adds, its increment, and measure how much of it an unaligned student recovers. The student fills the teacher's hierarchy from the bottom up and stalls near the top, which we call hierarchy filling: even after 400K steps it recovers almost none of the deepest. The recoverability gap is the unrecovered share of an increment, read from one unaligned checkpoint. In short runs that each align one teacher layer at one block, the gap nearly reproduces their ranking by FID improvement, and CKA, a measure of feature similarity, largely reverses it. Representation Alignment and Recoverability Estimation (RARE) picks the teacher layer with the largest gap before training. During training, it tracks each token's remaining distance to that layer, the online counterpart of the gap, weights tokens by it, and phases out the loss once the average distance stops falling. With SiT-B/2 on ImageNet 256×256256\times256, RARE reaches an FID of 18.02 without guidance and 4.46 with it, ahead of seven alignment baselines including REPA, iREPA and HASTE. It also trains in 14% fewer GPU-hours than iREPA. Its FID stays below iREPA's across model scales, teachers, datasets and backbones.
Oct 14, 2026

Representation Conditioning for Generation

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 head (as a proxy to the vision encoder) and project its output tokens to the later layers. Our method adds no parameters and leads to much faster convergence: RECON is 63×63\times faster than SiT and 25×25\times faster than REPA. Code is available at https://github.com/saeidrazavi/RECON
May 5, 2026

AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers

Representation alignment has recently emerged as an effective paradigm for accelerating Diffusion Transformer training. Despite their success, existing alignment methods typically impose a fixed supervision target or a fixed alignment granularity throughout the entire denoising trajectory, whether the guidance is provided by external vision encoders, internal self-representations, or VAE-derived features. We argue that such timestep-agnostic alignment is suboptimal because the useful granularity of representation supervision changes systematically with the signal-to-noise ratio. In high-noise regimes, diffusion models benefit more from coarse semantic and layout-level anchoring, whereas in low-noise regimes, the training signal should emphasize spatially detailed and structurally faithful refinement. This non-stationary alignment behavior creates a representational mismatch for static single-level supervisors. To address this issue, we propose Adaptive Hierarchical Prior Alignment (AHPA), a lightweight alignment framework that exploits the hierarchical representations naturally embedded in the frozen VAE encoder. Instead of using only a single compressed latent as the alignment target, AHPA extracts multi-level VAE features that provide complementary priors ranging from local geometry and spatial topology to coarse semantic layout. A timestep-conditioned Dynamic Router adaptively selects and weights these hierarchical priors along the denoising trajectory, thereby synchronizing the alignment granularity with the model's evolving training needs. Extensive experiments show that AHPA improves convergence and generation quality over baselines and incurs no additional inference cost while avoiding external encoder supervision during training.