A-SelecT: Automatic Timestep Selection for Diffusion Transformer Representation Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Liu, Changyu, Liang, James Chenhao, Yang, Wenhao, Cui, Yiming, Yang, Jinghao, Wang, Tianyang, Wang, Qifan, Liu, Dongfang, Han, Cheng
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912984172855296
author Liu, Changyu
Liang, James Chenhao
Yang, Wenhao
Cui, Yiming
Yang, Jinghao
Wang, Tianyang
Wang, Qifan
Liu, Dongfang
Han, Cheng
author_facet Liu, Changyu
Liang, James Chenhao
Yang, Wenhao
Cui, Yiming
Yang, Jinghao
Wang, Tianyang
Wang, Qifan
Liu, Dongfang
Han, Cheng
contents Diffusion models have significantly reshaped the field of generative artificial intelligence and are now increasingly explored for their capacity in discriminative representation learning. Diffusion Transformer (DiT) has recently gained attention as a promising alternative to conventional U-Net-based diffusion models, demonstrating a promising avenue for downstream discriminative tasks via generative pre-training. However, its current training efficiency and representational capacity remain largely constrained due to the inadequate timestep searching and insufficient exploitation of DiT-specific feature representations. In light of this view, we introduce Automatically Selected Timestep (A-SelecT) that dynamically pinpoints DiT's most information-rich timestep from the selected transformer feature in a single run, eliminating the need for both computationally intensive exhaustive timestep searching and suboptimal discriminative feature selection. Extensive experiments on classification and segmentation benchmarks demonstrate that DiT, empowered by A-SelecT, surpasses all prior diffusion-based attempts efficiently and effectively.
format Preprint
id arxiv_https___arxiv_org_abs_2603_25758
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle A-SelecT: Automatic Timestep Selection for Diffusion Transformer Representation Learning
Liu, Changyu
Liang, James Chenhao
Yang, Wenhao
Cui, Yiming
Yang, Jinghao
Wang, Tianyang
Wang, Qifan
Liu, Dongfang
Han, Cheng
Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
Image and Video Processing
Diffusion models have significantly reshaped the field of generative artificial intelligence and are now increasingly explored for their capacity in discriminative representation learning. Diffusion Transformer (DiT) has recently gained attention as a promising alternative to conventional U-Net-based diffusion models, demonstrating a promising avenue for downstream discriminative tasks via generative pre-training. However, its current training efficiency and representational capacity remain largely constrained due to the inadequate timestep searching and insufficient exploitation of DiT-specific feature representations. In light of this view, we introduce Automatically Selected Timestep (A-SelecT) that dynamically pinpoints DiT's most information-rich timestep from the selected transformer feature in a single run, eliminating the need for both computationally intensive exhaustive timestep searching and suboptimal discriminative feature selection. Extensive experiments on classification and segmentation benchmarks demonstrate that DiT, empowered by A-SelecT, surpasses all prior diffusion-based attempts efficiently and effectively.
title A-SelecT: Automatic Timestep Selection for Diffusion Transformer Representation Learning
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
Image and Video Processing
url https://arxiv.org/abs/2603.25758