Noise2Flow: trust the denoiser's direction, then walk 15% past its output

maier_ak · x · 2026-11-16

The author explains Noise2Flow's core observation: an imperfect denoiser is often right about the direction but wrong about the distance—the step from a noisy image toward its output usually points the right way (verified against ground-truth clean images). The method learns that direction from a single noisy image, fits a step length, and walks roughly 15% past the imperfect result, requiring only one network pass at the end—no clean reference or training set needed.

Related event: Noise2Flow Outperforms Prior Methods in Microscopy and Low-Dose CT(2 posts)→

Original post →

More from Research

Research channel →