Researchers cut MRI scan synthesis time from minutes to seconds using wavelet-based AI, enabling instant multi-scan generation on a single lightweight model

A team published WFM (Wavelet Flow Matching), a new AI method that generates four types of MRI brain scans (T1, T1c, T2, FLAIR) in just 1–2 computational steps instead of the 100–300 steps required by current AI systems, cutting processing time from minutes to seconds.

Unlike traditional AI image generators that start from random noise, WFM begins with the structural information already present in existing MRI scans (converted to wavelet format—a way of breaking down images into building blocks), then refines only the visual contrast. One 82-million-parameter model replaces four separate models that previously required 326 million parameters total.

Radiologists and hospitals can now synthesize missing or damaged MRI scans instantly at the point of care instead of waiting for lengthy processing, enabling faster diagnosis and reducing the need to reschedule patients for repeat scans when one modality fails or is unavailable.

The code and results are published on arXiv (open access); no commercial availability or release date has been announced yet, but the approach is designed for clinical adoption given its speed and single-model architecture.