New dual-stream AI sharpens electron microscopy images with better realism

A team led by Longmi Gao has introduced a frequency-aware dual-stream learning method to solve the long-standing trade-off between imaging resolution and acquisition speed in electron microscopy. Traditional single-stream AI models either oversmooth fine details (hurting realism) or generate artificial hallucinations (hurting fidelity). The new architecture first applies a discrete wavelet transform to decompose EM images into low-frequency structural components and high-frequency detail components.

Then, each stream is handled by a specialized model: a conditional diffusion model generates realistic global structures, while a transformer network recovers precise high-frequency details. Tested on the EMDiffuse dataset, the approach significantly outperforms existing methods in LPIPS (perceptual similarity) and resolution ratio. The code and dataset are publicly available, and the method generalizes well across diverse biological samples, supporting faster and more accurate nanoscale imaging for structural biology and nanotechnology applications.