Diffusion models have become a common tool for image super-resolution because their generative priors can produce sharp, perceptually pleasing results. But when the image contains text, those same priors can invent characters that were not in the original, a failure mode known as text hallucination.
A new arXiv paper, TGSR (Text-Grounded Diffusion for Faithful Image Super-Resolution), proposes to counter this by grounding the diffusion process in text. The approach is aimed specifically at keeping textual content faithful during upscaling, rather than letting the model freely improvise.
The paper's abstract notes the tension: diffusion priors boost perceptual quality, yet for textual content they often produce hallucinations. TGSR is positioned as a way to get the perceptual benefits of diffusion while reducing that risk for text.