Super-resolution models produce detailed images but often introduce unnatural patterns and texture distortions. Existing artifact detectors treat every defect as equally important, even though some are barely visible while others are highly distracting. To address this, the authors of a new arXiv paper propose artifact prominence: the fraction of viewers who judge a highlighted region to contain a noticeable artifact.

The researchers built SR-Prominence, a dataset suite with 3,935 artifact masks from DeSRA, Open Images, Urban100, and a no-ground-truth Urban100-HR setting, all annotated via a crowdsourced protocol. When they re-annotated DeSRA, they found that 48.2% of its in-lab binary artifacts were not noticed by a majority of viewers, suggesting that traditional labels overstate the perceptual impact of many defects.

Auditing existing tools, the team found that classical full-reference metrics, especially SSIM and DISTS, provided surprisingly strong localized prominence signals. In contrast, no-reference image-quality assessment methods and specialized artifact detectors often failed to generalize across datasets and reference settings. The suite is released with an objective scoring protocol, so new metrics can be benchmarked without additional crowdsourcing, moving super-resolution artifact evaluation from binary presence toward perceptual impact. The paper was accepted to the NeurIPS 2026 Evaluations and Datasets track.