OcuBench and OcuFlow: A New Paradigm for Tackling Eyeglass Reflection Removal

Eyeglass reflections are a long-overlooked yet notoriously difficult problem in facial image processing. Whether in smartphone selfies, video conferencing, or any face-centric visual application, lens reflections can range from subtle photometric noise to severe ocular occlusion. A new paper published on arXiv introduces OcuBench, a multi-source benchmark dataset, and OcuFlow, an ocular-adaptive pixel-level MeanFlow framework, aiming to systematically address this challenge.

Why Eyeglass Reflection Removal Is So Hard

The complexity of eyeglass reflections lies in their inherent "uncertainty." Reflection intensity spans an enormous range — sometimes it's just a faint glare on the lens that barely affects visibility; other times it completely obscures the eyes, causing severe ocular occlusion. This means algorithms cannot apply a one-size-fits-all approach across all regions, nor can they compromise a person's identity or natural appearance during restoration.

In other words, ideal reflection removal requires "selective correction": reconstructing only the regions genuinely contaminated by reflections while preserving reliable original details elsewhere in the image. This places demanding requirements on both a model's region-awareness and reconstruction quality.

OcuBench: A Multi-Source Benchmark That Fills the Dataset Gap

The research team notes that existing datasets cover only a limited range of reflection conditions, making it difficult to support model generalization in complex real-world scenarios or enable systematic evaluation. To address this, they constructed OcuBench as a multi-source benchmark.

Its data composition is thoughtfully layered:

• 10,280 controlled synthetic image pairs: for paired evaluation under controlled conditions, covering diverse reflection scenarios;
• 732 real-input pseudo-pairs: pseudo-pairs constructed from real-world captured data;
• 458 standalone real-world test images: supporting evaluation beyond generative supervision.

This three-tier design of "synthetic + pseudo-pairs + real-world testing" balances the rigor of paired evaluation with the diversity of real-world scenarios, enabling assessments that are both quantifiable and practically relevant. This is precisely the systematic evaluation foundation that previous research has lacked.

OcuFlow: A One-Step Ocular-Adaptive Reconstruction Framework

On the methodology side, the research introduces OcuFlow — an ocular-adaptive pixel-level MeanFlow (pMF) framework whose core goal is efficient, detail-preserving image restoration.

Geometry-Adaptive Representation and One-Step pMF

OcuFlow combines geometry-adaptive representation with one-step pMF. The design philosophy is to focus the reconstruction process on the eye regions occluded by reflections, rather than processing the entire image indiscriminately. This targeted strategy improves restoration precision while avoiding unintended alterations to uncontaminated regions.

Geometry-adaptive representation refers to the model's ability to dynamically adjust its processing based on the shape and spatial position of specific regions in the input image, rather than applying a uniform receptive field or fixed convolution kernel across the entire image. In eyeglass reflection scenarios, reflection boundaries are often irregular, varying widely due to head pose, frame shape, and light source position. The geometry-adaptive design enables the network to perceive and align these irregular occlusion contours, precisely concentrating restoration effort on the contaminated eye regions. This shares some design philosophy with Partial Convolution or Deformable Convolution in image inpainting, but in OcuFlow it is further integrated with pMF's one-step generation mechanism to form an architecture specifically tailored for reflection removal.

Native-Resolution Frequency-Preserving Synthesis

Another key design is "native-resolution frequency-preserving synthesis." By preserving frequency information at the original resolution, the model retains reliable observed details from the image, removing reflections without sacrificing visual authenticity or sharpness. This is especially important for facial images — any loss of detail can affect identity recognition and natural appearance.

MeanFlow is a generative modeling paradigm based on Flow Matching, which learns a continuous velocity field from a noise distribution to a data distribution for image generation or restoration. Unlike diffusion models that require tens or even hundreds of denoising inference steps, the key innovation of MeanFlow lies in parameterizing the average velocity of an entire trajectory as a single network output, theoretically enabling one-step forward inference for high-quality image synthesis. Pixel MeanFlow (pMF) implements this mechanism in pixel space (rather than latent space), avoiding the additional information loss introduced by encoders and decoders — particularly well-suited for reflection removal tasks that require preserving fine textures and ocular details. The one-step inference characteristic also significantly reduces computational overhead, giving OcuFlow practical deployment efficiency alongside high-quality output.

The concept of frequency-preserving synthesis originates from frequency-domain analysis in image signal processing. Images can be decomposed into low-frequency components (large-scale structures, overall tone) and high-frequency components (edges, textures, details). Many deep learning image generation models tend to lose high-frequency details in encoder-decoder architectures with downsampling and upsampling (such as U-Net), leading to overly smooth outputs or blurring artifacts. In facial images, eyelashes, iris textures, and eyeglass frame edges are precisely high-frequency information — once lost, restoration results appear unnatural and can impair downstream face recognition applications. OcuFlow operates at native resolution and explicitly maintains these details through frequency-preserving mechanisms, ensuring that restored eye regions remain consistent with surrounding undisturbed image areas in terms of frequency characteristics, thereby enhancing overall visual realism.

Experimental Results: Triple Advantages in Quality, Fidelity, and Efficiency

The research conducted experiments across various reflection conditions, and results show that OcuFlow demonstrates consistent advantages across three dimensions:

• Reflection removal quality: effectively eliminates reflections ranging from mild to severe;
• Ocular fidelity: preserves the natural and authentic appearance of the eye region while restoring occluded areas;
• Efficiency: the one-step pMF design delivers computational efficiency.

The most compelling evidence comes from a blind user study: OcuFlow achieved a selection rate of 67.32%, representing 6.2 times the share of the second-best method. This overwhelming subjective preference indicates that its output not only excels on objective metrics but also holds a clear advantage in human perceptual terms.

Open Resources and Potential Impact

The research team has stated that both code and datasets will be publicly released. This means OcuBench has the potential to become a universal evaluation standard for this niche domain, while OcuFlow's technical approach provides a reproducible reference for future research.

From an application perspective, eyeglass reflection removal directly affects selfie enhancement, video conferencing image quality improvement, face recognition preprocessing, and more. As more people wear glasses, the practical value of such technology will continue to grow. The combination of OcuBench and OcuFlow provides both a "ruler" for evaluation and a "tool" for solving the problem, laying the foundation for research and real-world deployment in this direction.

Conclusion

The significance of this work lies not only in a higher-performing algorithm, but also in how it redefines the evaluation paradigm for eyeglass reflection removal through a multi-source benchmark. When selective correction, detail preservation, and computational efficiency are simultaneously taken into account, image restoration moves from "can remove" to "removes correctly and authentically." With the open release of code and data, this field is poised for faster iteration.