OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution
Shubhashis Roy Dipta, Sourajit Saha, Shaswati Saha, Nobin Sarwar
University of Maryland, Baltimore County
Published September 6, 2026 · on the feed September 10, 2026 · ▲ 7 · arXiv 2609.06490
The abstract
Recursive Super-Resolution (SR) extends fixed-scale SR to extreme magnification by repeatedly feeding predictions back into the same model, analogous to zooming an image repeatedly. However, ground truth availability at every scale, especially at depth, remains challenging as the required source resolution grows geometrically, leaving deeper predictions unsupervised. We present OracleZoom, an on-policy distillation-inspired, reference-constrained framework that trains on its trajectory while carrying the last ground-truth evidence beyond the supervision boundary. Direct and cross-scale supervision constrain verifiable content, while a no-reference quality objective guides unresolved fine-scale detail. A KL-constrained pretrained latent prior limits quality-driven drift, while EMA consistency stabilizes the supervision boundary. Across seven datasets, OracleZoom achieves the state-of-the-art SR quality across zooming scales, averaging 0.713 CLIPIQA, with larger gains on deeper scales, while significantly reducing hallucinations. Code, data, and models are available at https://dipta007.github.io/OracleZoom/ .
The Brief
1:56A short podcast summary of the paper. Written and voiced by AI.
Transcript
OracleZoom. On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution. By Shubhashis Roy Dipta and three others, from the University of Maryland, Baltimore County. Published on the sixth of September. Its code repository is worked out: about forty source files, with a readme and a licence.
Recursive super-resolution pushes image upscaling to extreme magnification by feeding a model's prediction back into the same model, again and again, much like zooming into an image repeatedly. The trouble is supervision. Ground truth has to exist at every scale, and the source resolution needed for it grows geometrically with depth. Past a certain point there simply is none, and the deeper predictions go unsupervised.
OracleZoom is a reference-constrained framework inspired by on-policy distillation. It trains on its own zoom trajectory, and it carries the last available ground-truth evidence forward, beyond the boundary where direct supervision stops.
Several constraints divide the work. Direct supervision and cross-scale supervision constrain the content that can be verified. For fine detail that cannot be resolved against any reference, a no-reference quality objective provides the guidance. To stop that quality objective from pulling the output astray, a pretrained latent prior with a KL constraint limits the drift. And consistency through an EMA keeps the supervision boundary itself stable.
Across seven datasets, OracleZoom reaches state-of-the-art super-resolution quality across zoom scales, with an average CLIPIQA of zero point seven one three. The gains are larger at the deeper scales, and hallucinations are significantly reduced. Code, data and models are available.
Carrying the last real evidence forward, rather than abandoning it, is the core of the design. Beyond the boundary, the model is guided by what could still be verified, by a quality signal, and by a prior that keeps it honest, which is where the reduced hallucination at depth is reported.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Zoom in far enough and the truth runs out. What then?
You upscale an image, then feed the result back into the same model to zoom in again, and again. Each step goes deeper into detail.
The road ahead
- 1 · The problem ~7 min Zooming an image by feeding each upscaled result back into the same model soon runs out of real images to train against.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Train on the model's own zoom chain, check deep zooms against the last real image, and let a held-back quality score fill the rest.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Run the zoom chain on the model's own outputs, then apply different losses where targets exist, just past them, and deeper still.
- 6 · The experiment ~6 min You’ll guess: Across seven datasets, OracleZoom was compared with other super-resolution methods at several zoom depths. Where were its quality gains largest?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.