New A-SelecT method automatically identifies optimal timesteps in Diffusion Transformers, boosting efficiency for representation learning without exhaustive searching.

A-SelecT dynamically selects the most information-rich timestep from Diffusion Transformer features in a single run, eliminating computationally expensive exhaustive searches

Diffusion Transformers (DiT) are emerging as promising alternatives to U-Net-based diffusion models for both generative and discriminative tasks

The approach addresses current limitations in DiT training efficiency and representational capacity by better exploiting transformer-specific feature representations

Automatic timestep selection enables improved downstream discriminative task performance through more effective generative pre-training