Text-to-Image Diffusion Sample Selector for Learning With Noisy Labels

Large-scale datasets inevitably contain noisy labels, compelling deep neural networks to overfit and degrade in performance. Existing learning with noisy labels (LNL) approaches partially alleviate this through sample selection. However, selection processes themselves inherently rely on models trained on noisy data, which makes them prone to self-confirmation bias. To overcome this issue, we decouple sample selection from model training by proposing a Text-to-Image Diffusion Sample Selector (TIDS), which can provide a clean initial subset for downstream learning, even under various complex and extremely noisy conditions. Theoretically, we constrain the selection criteria by minimizing the generalization error bound in noisy environments to ensure the reliability of the selected samples and facilitate the effective training of robust models. Our proposed TIDS innovatively explores the semantic understanding capabilities of diffusion models. It estimates the label noise probability by evaluating the matching confidence between noisy textual labels and corresponding images, while capturing class information that more closely approximates the ground-truth. Therefore, TIDS enables a two-stage data cleaning process: correction and selection, both of which are single-step, generic preprocessing techniques that can work out of the box with most LNL methods to improve robustness. Furthermore, we design an efficient and lightweight training framework tailored to TIDS, achieving superior performance on both synthetic and real-world noisy datasets. Code for reproducing our experiments is available at https://github.com/SenyuHou/TIDS.