📄 中文摘要
扩散模型在生成人工智能领域中产生了重大影响，并逐渐被探索用于判别表示学习。扩散变换器（DiT）作为一种有前景的替代传统U-Net扩散模型的方法，展现了通过生成预训练在下游判别任务中的潜力。然而，其当前的训练效率和表示能力受到时间步搜索不足和DiT特定特征表示利用不足的限制。为了解决这一问题，提出了自动选择时间步（A-SelecT），该方法能够动态识别DiT中最具信息量的时间步，从而提升模型的性能和效率。
A-SelecT: 扩散变换器表示学习的自动时间步选择
📄 English Summary
A-SelecT: Automatic Timestep Selection for Diffusion Transformer Representation Learning
Diffusion models have significantly impacted the field of generative artificial intelligence and are increasingly being explored for their potential in discriminative representation learning. The Diffusion Transformer (DiT) has emerged as a promising alternative to conventional U-Net-based diffusion models, demonstrating potential for downstream discriminative tasks through generative pre-training. However, its current training efficiency and representational capacity are largely limited by inadequate timestep searching and insufficient exploitation of DiT-specific feature representations. To address this, Automatically Selected Timestep (A-SelecT) is introduced, which dynamically identifies the most informative timesteps in DiT, thereby enhancing the model's performance and efficiency.