Researchers have released OracleZoom, a framework for recursive image super-resolution (SR) that aims to keep extreme magnification grounded in real image evidence. The work appears as a paper on Hugging Face and addresses a core limitation of recursive SR: as predictions are repeatedly fed back into the same model, ground-truth data becomes unavailable at deeper scales because the required source resolution grows geometrically.
OracleZoom is described as an on-policy distillation-inspired, reference-constrained approach. It trains on its own trajectory while carrying the last available ground-truth evidence beyond the supervision boundary, so that deeper predictions are not left entirely unsupervised.
The method combines several objectives: direct and cross-scale supervision to constrain verifiable content, a no-reference quality objective to guide unresolved fine-scale detail, a KL-constrained pretrained latent prior to limit quality-driven drift, and EMA consistency to stabilize the supervision boundary.
Across seven datasets, the authors report state-of-the-art SR quality across zooming scales, with an average of 0.713 CLIPIQA and larger gains at deeper scales, alongside a significant reduction in hallucinations. Code, data, and models are said to be available via the project page.