| Author | Kseniia Ibragimova |
| Supervisor | Andrei Filatov |
This work investigates the applicability of the Generative Drifting framework to one-step image super-resolution in the one-to-many setting. A single low-resolution image may correspond to multiple plausible high-resolution reconstructions, yet most efficient super-resolution methods are trained deterministically and therefore tend to produce averaged, over-smoothed outputs. To address this limitation, we explore the Generative Drifting paradigm, which models generation through a learned drift field and enables single-step inference. We propose a conditional stochastic formulation that learns a direct mapping from a low-resolution image and noise to plausible high-resolution images in a single forward pass. The proposed method defines drifting in residual feature space, encouraging generated samples to move toward plausible high-resolution targets while preventing collapse to a single solution. Experiments evaluate reconstruction quality, perceptual realism, and diversity using PSNR, SSIM, LPIPS, and one-to-many consistency metrics.
Generative Drifting, Super-Resolution, One-Step Generation
Our project is MIT licensed. See LICENSE for details.