Where do AI image upscalers actually fail?

Diffusion-based image upscalers produce sharp results but introduce subtle, hard-to-describe artifacts. SR-Ground, a 63,000-image dataset with pixel-level annotations for 6 artifact types (created via crowdsourced validation from 1,062 people), lets researchers train quality models that explain *why* an upscale looks wrong instead of just scoring it. Models trained on this grounding data improve downstream artifact detection and can be fine-tuned to reduce visible defects in upscaler outputs.