Generative image super-resolution aims to produce visually realistic high-resolution images, but it can also introduce hallucinations—adding unsupported content or altering existing details. This is a known risk when models fill in missing information during upscaling.
Existing super-resolution metrics are not designed to detect such hallucinations, leaving a gap in evaluation. The new approach, VHA (VQA-based Hallucination Assessment), proposes using visual question answering to assess whether generated details are faithful to the source image.
The abstract notes that existing metrics fail, but it does not detail the specific methodology or results. Still, the proposal points toward a more robust way to evaluate generative super-resolution models, particularly for applications where fidelity matters.