EchoAlign: Bridging Generative and Discriminative Learning under Noisy Labels

EchoAlign learns from noisy labels by editing images with controllable generators and selecting clean originals via feature similarity, greatly improving robustness to instance-dependent noise.