MeanFlow for Image Inverse Problems
One-step image denoising by starting MeanFlow from noisy images instead of Gaussian noise.
Undergraduate research at ZJUI, advised by Prof. Xiangming Meng, Jun 2025 – Dec 2025.
MeanFlow models learn the average velocity over a time interval, not the instantaneous velocity. This lets them generate a sample in a single step. In this project I changed the initial distribution of MeanFlow from a Gaussian to the distribution of noisy images. The model then learns a direct one-step map from a corrupted image to a clean one, which gives efficient one-step denoising.