OcuFlow Removes Eyeglass Glare in One Step and Wins 67.32% of Votes in a Blind Test

A team of authors on arXiv has presented two artifacts for one narrow but very practical task: removing glare from eyeglasses in portraits. The first is the OcuBench benchmark, with 10 280 controllable synthetic pairs, 732 pseudo-pairs on real inputs and 458 independent real test images. The second is the OcuFlow model, which restores the eye region in a single generation step. In a blind user study it received 67.32% of the choices, 6.2 times more than the next-best method. The authors promise to publish the code and dataset.

Why glare on eyeglasses is such a problem

The task looks minor until you consider where it comes up. Smartphone cameras, video calls, selfie filters, any app that relies on the face: wherever a person wearing glasses sits in front of a window or a screen, stray glare appears on the lenses. Sometimes it is a slight shift in color and contrast. Sometimes the lens covers the eye completely, and there is simply no information under the glare.

This creates a dual requirement. Weak glare needs a targeted correction with no side effects. Strong glare needs a plausible reconstruction that does not change the person's identity and does not give the skin and iris a "plastic" look. The authors point to a weakness in existing datasets: they cover a narrow range of lighting conditions and reflections, so models transfer worse to complex real-world scenes, and comparisons between methods end up incomplete.

How OcuBench and OcuFlow work

OcuBench consists of three parts, each with a different role. The synthetic pairs provide controlled ground truth: it is known what the frame looked like without the glare. The pseudo-pairs on real inputs bring the distribution closer to real life. The independent real images have no generated reference at all, and they are used to check whether the model has merely learned to guess synthetic artifacts.

OcuFlow is built on pixel MeanFlow (pMF), a one-step generative approach that works in pixel space. Instead of multi-step denoising, as in classic diffusion, the model performs the restoration in a single pass. For video calls this matters, because latency is critical there.

According to the authors, the architecture rests on two key decisions:

• a geometry-adaptive representation directs the reconstruction to the eye regions covered by glare, rather than to the whole face;
• native-resolution synthesis with frequency preservation carries reliable detail from the original frame into the result, so that the texture of skin, eyelashes and frames is not blurred.

What is known about the results

The main reported figures:

• 10 280 synthetic pairs, 732 real pseudo-pairs and 458 independent real images in OcuBench;
• a consistent OcuFlow advantage in glare-removal quality, eye-region fidelity and efficiency across different types of reflections;
• 67.32% of choices in the blind user survey;
• a 6.2× lead over second place: from these figures, the nearest competitor has roughly 10.9% of the votes.

Two points call for caution. First, the abstract contains no numeric metrics such as PSNR or LPIPS and no speed data in milliseconds, so the efficiency claim can only be verified against the full text and the code. Second, a blind survey measures subjective preference but does not guarantee that facial identity is preserved in every case. For eye reconstruction under heavy glare this is a fundamental risk: the model paints in what it has not seen.

Who can use this, and how

If the authors keep their promise and publish the code and data, the industry will gain a shared testbed for comparing such methods, which usually moves a field forward faster than individual models do. For developers of video-calling software and smartphone cameras, the one-step mode is the point of interest: it is easier to build into a real-time pipeline than multi-step diffusion.

The practical plan is simple. Wait for the release and run OcuFlow on your own frames with heavy glare, not only on OcuBench. Check identity preservation separately by comparing face embeddings before and after processing. Measure latency on the target hardware. And decide in advance how to label frames in which the eye region was reconstructed rather than cleaned: for identity verification and for documents, this is critical.