REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training

• URL: http://arxiv.org/abs/2505.16792v1
• Date: Thu, 22 May 2025 15:34:33 GMT
• Title: REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training
• Authors: Ziqiao Wang, Wangbo Zhao, Yuhao Zhou, Zekai Li, Zhiyuan Liang, Mingjia Shi, Xuanlei Zhao, Pengfei Zhou, Kaipeng Zhang, Zhangyang Wang, Kai Wang, Yang You,
• Abstract summary: Diffusion Transformers (DiTs) deliver state-of-the-art image quality, yet their training remains notoriously slow.<n>A recent remedy -- representation alignment (REPA) that matches DiT hidden features to those of a non-generative teacher (e.g. DINO) -- dramatically accelerates the early epochs but plateaus or even degrades performance later.<n>We trace this failure to a capacity mismatch: once the generative student begins modelling the joint data distribution, the teacher's lower-dimensional embeddings and attention patterns become a straitjacket rather than a guide.<n>We introduce HASTE
• Score: 58.33728862521732
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Diffusion Transformers (DiTs) deliver state-of-the-art image quality, yet their training remains notoriously slow. A recent remedy -- representation alignment (REPA) that matches DiT hidden features to those of a non-generative teacher (e.g. DINO) -- dramatically accelerates the early epochs but plateaus or even degrades performance later. We trace this failure to a capacity mismatch: once the generative student begins modelling the joint data distribution, the teacher's lower-dimensional embeddings and attention patterns become a straitjacket rather than a guide. We then introduce HASTE (Holistic Alignment with Stage-wise Termination for Efficient training), a two-phase schedule that keeps the help and drops the hindrance. Phase I applies a holistic alignment loss that simultaneously distills attention maps (relational priors) and feature projections (semantic anchors) from the teacher into mid-level layers of the DiT, yielding rapid convergence. Phase II then performs one-shot termination that deactivates the alignment loss, once a simple trigger such as a fixed iteration is hit, freeing the DiT to focus on denoising and exploit its generative capacity. HASTE speeds up training of diverse DiTs without architecture changes. On ImageNet 256X256, it reaches the vanilla SiT-XL/2 baseline FID in 50 epochs and matches REPA's best FID in 500 epochs, amounting to a 28X reduction in optimization steps. HASTE also improves text-to-image DiTs on MS-COCO, demonstrating to be a simple yet principled recipe for efficient diffusion training across various tasks. Our code is available at https://github.com/NUS-HPC-AI-Lab/HASTE .

Related papers

• Representation Entanglement for Generation:Training Diffusion Transformers Is Much Easier Than You Think [56.539823627694304]
  REPA and its variants effectively mitigate training challenges in diffusion models by incorporating external visual representations from pretrained models.<n>We argue that the external alignment, which is absent during the entire denoising inference process, falls short of fully harnessing the potential of discriminative representations.<n>We propose Representation Entanglement for Generation (REG), which entangles low-level image latents with a single high-level class token from pretrained foundation models for denoising.
  arXiv  Detail & Related papers  (2025-07-02T08:29:18Z)
• Autoregressive Distillation of Diffusion Transformers [18.19070958829772]
  We propose AutoRegressive Distillation (ARD), a novel approach that leverages the historical trajectory of the ODE to predict future steps.<n>ARD offers two key benefits: 1) it mitigates exposure bias by utilizing a predicted historical trajectory that is less susceptible to accumulated errors, and 2) it leverages the previous history of the ODE trajectory as a more effective source of coarse-grained information.<n>Our model achieves a $5times$ reduction in FID degradation compared to the baseline methods while requiring only 1.1% extra FLOPs on ImageNet-256.
  arXiv  Detail & Related papers  (2025-04-15T15:33:49Z)
• Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards [52.90573877727541]
  reinforcement learning (RL) has been considered for diffusion model fine-tuning.<n>RL's effectiveness is limited by the challenge of sparse reward.<n>$textB2text-DiffuRL$ is compatible with existing optimization algorithms.
  arXiv  Detail & Related papers  (2025-03-14T09:45:19Z)
• Improved Distribution Matching Distillation for Fast Image Synthesis [54.72356560597428]
  We introduce DMD2, a set of techniques that lift this limitation and improve DMD training. First, we eliminate the regression loss and the need for expensive dataset construction. Second, we integrate a GAN loss into the distillation procedure, discriminating between generated samples and real images.
  arXiv  Detail & Related papers  (2024-05-23T17:59:49Z)
• SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer [102.39050180060913]
  Diffusion Transformer (DiT) has emerged as the new trend of generative diffusion models on image generation. Recent breakthroughs have been driven by mask strategy that significantly improves the training efficiency of DiT with additional intra-image contextual learning. In this work, we address these limitations by novelly unleashing the self-supervised discrimination knowledge to boost DiT training.
  arXiv  Detail & Related papers  (2024-03-25T17:59:35Z)
• You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs [13.133574069588896]
  YOSO is a novel generative model designed for rapid, scalable, and high-fidelity one-step image synthesis with high training stability and mode coverage.<n>We show that our method can serve as a one-step generation model training from scratch with competitive performance.<n>In particular, we show that the YOSO-PixArt-$alpha$ can generate images in one step trained on 512 resolution, with the capability of adapting to 1024 resolution without extra explicit training, requiring only 10 A800 days for fine-tuning.
  arXiv  Detail & Related papers  (2024-03-19T17:34:27Z)
• ACT-Diffusion: Efficient Adversarial Consistency Training for One-step Diffusion Models [59.90959789767886]
  We show that optimizing consistency training loss minimizes the Wasserstein distance between target and generated distributions. By incorporating a discriminator into the consistency training framework, our method achieves improved FID scores on CIFAR10 and ImageNet 64$times$64 and LSUN Cat 256$times$256 datasets.
  arXiv  Detail & Related papers  (2023-11-23T16:49:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.