Researchers propose NR-CFM to clean noisy datasets

A team of researchers from the University of Warsaw and ETH Zurich has introduced Noise-Robust Conditional Flow Matching (NR-CFM), a novel approach to generate clean samples from noisy datasets. Published on arXiv under the title 'Noise-Robust Conditional Flow Matching: Generating Clean Samples from Noisy Datasets,' the method addresses a critical challenge in generative modeling: the need for high-quality training data.

NR-CFM leverages conditional flow matching (CFM), a framework known for its stable training, efficient sampling, and strong image-generation performance. The researchers extended CFM to handle noisy data by providing a closed-form clean endpoint correction for additive white Gaussian noise and learning a data-driven correction for more complex Gaussian corruptions. In evaluations, NR-CFM outperformed NR-GAN in most cases and remained competitive with Ambient Diffusion, particularly in high-noise regimes. The team demonstrated its effectiveness on scientific data with signal-to-noise ratios as low as 0.001, generating plausible particle images from severely corrupted measurements.