Seeing Through the Glare: A Multi-Source Benchmark and Ocular-Adaptive Pixel MeanFlow for Eyeglass Reflection Removal

• Reflections can shift color or severely occlude ocular structures, so a portrait may need correction, reconstruction, and preservation in different regions.Removing reflections without unnecessary portrait changes requires accommodating these different restoration needs.Source: §1 · p. 1
• Existing data have correspondence problems in captured pairs, cannot establish natural-reflection performance through synthetic pairs alone, and cannot verify obscured appearance in real-input pseudo-pairs.Reliable assessment therefore requires complementary sources and real-world evaluation independent of generated targets.Source: §1 · pp. 1–2
• Uniform tokenization can spend limited computation on unaffected regions, while reflection strength alone cannot indicate how much ocular structure remains visible.Compact patches can also discard native details, whereas preserving high-frequency input content indiscriminately risks retaining reflection artifacts.Source: §1 · p. 2
• OcuBench combines 10,280 synthetic pairs, 732 real-input pseudo-pairs, and 458 independent real-world test images for controlled and real-world evaluation.OcuFlow uses geometry-adaptive supports under a fixed token budget, one-pass patch restoration, and native-resolution frequency-preserving synthesis.Source: §1 · pp. 2–3
• OcuFlow achieves superior overall restoration quality on the held-out paired benchmark and runs in 56 ms per image, with 4–66× speedups over compared methods.It also achieves the highest user preference and best both-eyes localization accuracy among compared methods.Source: §1 · p. 3

• Existing reflection datasets and methods range from synthetic overlays and generated-portrait pairs to filtered real pairs, constrained optimization, supervised restoration, and generative reconstruction.Synthetic realism and real-pair alignment remain challenges.Source: §2 · p. 3
• Adaptive representations reduce or vary visual tokens, while OcuFlow uses ocular geometry to allocate a fixed token budget to fine eye structures and coarser regions elsewhere.Its tokens retain position and scale for geometry-guided restoration and native-resolution reconstruction.Source: §2 · p. 3
• Generative restoration ranges from iterative diffusion and flow methods to single-pass priors and one-step reflection removal, while OcuFlow adapts pMF to ocular patches.Its renderer blends predictions with native-resolution observations by frequency to selectively retain high-frequency details beyond compact patches.Source: §2 · p. 3

3 OcuBench: A Multi-Source Benchmark

• OcuBench combines synthetic pairs, real-input pseudo-pairs, and independent real-world samples because real paired facial data are difficult to collect and synthetic data alone cannot capture natural reflections fully.The sources support controlled degradation analysis, realistic reflection coverage, and independent real-world evaluation.Source: §3 · p. 4, Figure 2 · p. 5

3.1 Data Collection and Generation

• Ocu-S hierarchically samples subject, imaging, and reflection factors, while Ocu-P preserves naturally reflected inputs and generates only clean targets.Ocu-R combines captured reference pairs and unpaired images from unconstrained scenes; its reflected inputs and references are acquired without image generation.Source: §3.1 · p. 4, Figure 2 · p. 5

3.2 Data Cleaning and Quality Control

• Reviewers and an edge-consistency filter assess pair quality; the retained datasets contain 10,280 Ocu-S samples and 732 Ocu-P samples.Pseudo-clean supervision remains approximate because obscured ocular appearance cannot be verified from reflected inputs alone.Source: §3.2 · p. 5

3.3 Dataset Analysis

• Ocu-S samples balance four requested ocular-occlusion levels, while Ocu-R reaches substantially higher resolutions than Ocu-S and Ocu-P.The sources also differ in luminance, aspect ratio, high-frequency detail, and saturation, motivating separate evaluation on independently captured images.Source: §3.3 · p. 6

4 Method

• OcuFlow estimates a reflection-free output at the input resolution by allocating fixed patch capacity using ocular geometry, restoring patches in one pass, and fusing predictions with native-resolution observations.Reflected observations, ocular geometry, and coarse context condition the restoration.Source: §4 · p. 6

4.1 Ocular-Prior-Guided Adaptive Representation

• A fixed-budget quadtree partition assigns finer patches to ocular regions and coarser patches to largely unaffected regions.Landmark-derived ocular priors guide partitioning; under a 256-token budget, higher-importance regions receive smaller leaves.Source: §4.1 · p. 7

4.2 Geometry-Guided One-Step Pixel MeanFlow

• Geometry-guided pMF conditions patch restoration on reflected observations and ocular context, producing restored patches in one network evaluation at inference.The observation-anchored source is formed by adding small ocular-prior-weighted Gaussian perturbations to reflected adaptive patches.Source: §4.2 · p. 7

4.3 Native-Resolution Frequency-Preserving Synthesis

• Native-resolution synthesis combines absolute patch reconstruction with a residual correction to the input, then selectively mixes frequency components to balance photographic detail and reflection artifacts.The difference between residual and absolute candidates represents information not preserved through patch sampling and projection; spatial maps control low- and high-frequency mixing within lens support.Source: §4.3 · p. 8

5.1 Metrics and Implementation Details

• Evaluation measures paired reconstruction, ocular similarity and geometry, real-world no-reference quality, and processing time under shared 1008 × 1296 paired-inference conditions.Paired metrics use original paired resolutions; Ocu-R is evaluated at its original resolution.Source: §5.1 · p. 8

5.2 Evaluation

• OcuFlow ranks first on six of seven paired restoration and ocular geometry metrics and runs 4.3× faster than ER2Net, the second-fastest method.The paired comparison uses the held-out OcuBench test set and a shared evaluation protocol.Source: §5.2 · p. 8, Table 2 · p. 9
• On 458 real-world images, OcuFlow receives 67.32% of valid user selections, ranks first on 26 of 30 images, and has the lowest NIQE among restoration methods.Its CLIP-IQA+ score is second-highest among restoration methods, while the reflected input scores higher, so no-reference metrics are treated as auxiliary indicators.Source: Figure 5 · p. 9, §5.2 · p. 9
• Qualitative comparisons show OcuFlow suppressing localized glare and broad blue reflections while recovering clearer iris and eyelid contours and preserving reliable facial details.Compared methods retain haze or color casts, soften eyes, or introduce irregular details around rims and eyelids.Source: §5.2 · p. 9, Figure 6 · p. 10

5.3 Ablation Study

• Ocular-adaptive allocation improves masked reconstruction, Eye-ROI LPIPS, and both ocular NME measures over a uniform grid, while outside-mask PSNR measures remain nearly unchanged.The uniform grid is faster, and removing geometry embedding reduces reconstruction and ocular fidelity.Source: §5.3 · pp. 9–10, Table 3 · p. 10
• Removing geometry-aware input or coarse-to-fine conditioning worsens masked reconstruction and all eye-specific metrics; 1-NFE outperforms 2-NFE on masked reconstruction and ocular geometry at lower runtime.Although 2-NFE has lower Eye-ROI LPIPS, its added computation does not consistently improve restoration.Source: §5.3 · pp. 10–11
• The frequency-preserving renderer performs best across its ablation block, improving outside-mask RGB and high-frequency agreement, while residual rendering improves masked reconstruction but slightly worsens ocular geometry.The results support selective frequency fusion to balance restoration and preservation.Source: §5.3 · p. 11
• Adding real-input pseudo-pairs improves paired PSNR/SSIM and Eye-NME and yields the highest real-world CLIP-IQA+, though Iris-NME rises slightly and synthetic-data scaling does not monotonically improve CLIP-IQA+.Increasing the synthetic fraction consistently improves paired reconstruction and ocular geometry.Source: §5.3 · p. 11, Table 4 · p. 11

6 Conclusion

• OcuBench combines synthetic pairs, real-input pseudo-pairs, and independent real-world evaluation, while OcuFlow uses ocular geometry, one-step restoration, and frequency-preserving synthesis to reconstruct obscured eyes and retain reliable details.The conclusion reports improved reflection removal, ocular reconstruction, downstream iris localization, human preference, and lower inference cost; future work could extend the approach to temporally consistent video restoration.Source: §6 · p. 11

A.1.1 Hierarchical Factor Sampling

• Ocu-S generation hierarchically samples subject, imaging, and reflection factors, conditioning later choices on earlier ones to span compatible reflection conditions.Target ocular-occlusion levels are approximately balanced across mild, moderate, strong, and severe cases, while reflection-strength levels are sampled non-uniformly.Source: §A.1.1 · p. 15

A.1.2 Pair Generation

• Synthetic-pair generation uses a fixed random seed and stored prompts, enabling interrupted runs to resume without resampling or duplicating factor combinations.All scheduled samples were generated; the full synthetic set took approximately 17 days.Source: §A.1.2 · pp. 16–17
• For each sampled factor tuple, structured prompts generate a clean portrait and then edit it to add a realistic lens reflection while preserving identity, face and ocular geometry, pose, eyeglasses, and composition.The reflection prompt constrains the change to the transparent lens regions and specifies reflection attributes and target occlusion.Source: §A.1.2 · p. 16

A.1.3 Pair Filtering

• Synthetic pairs are aligned using a similarity transform, assessed with regional edge distances that exclude reflection-affected pixels, automatically rejected above a 3-pixel threshold, and then manually reviewed.Pairs are rejected if reviewers find identity drift, deformation, pose or gaze changes, or unintended edits; 10,280 pairs are retained.Source: §A.1.3 · pp. 17–19

A.2.1 Real Reflected-input Collection

• Real reflected inputs are screened using CLIP ranking for SoF and FaceMesh-guided local glare analysis for Web images, followed by manual review.The resulting pseudo-pair set contains 199 SoF and 533 Web images.Source: §A.2.1 · pp. 19–20, Table 6 · p. 19, §A.2.3 · p. 20

A.2.2 Clean-target Generation

• Pseudo-pairs keep each original real reflected input unchanged and use a generated reflection-free image only as its clean target, so their reflection distribution comes directly from real observations.All 732 resulting pairs passed pair-level quality control and were retained.Source: §A.2.2 · p. 20, §A.2.3 · p. 20

A.3.1 Reference-pair Capture and Processing

• The paired-reference subset retains 157 pairs—62 offline and 95 online—after consistency checks and color correction; clean images serve as visual references, not pixel-level ground truth.Its reflected images are included in Ocu-R alongside 301 unpaired images.Source: §A.3.1 · p. 21

A.3.2 In-the-wild Test Images

• The 301 unpaired Ocu-R images were opportunistically captured in everyday settings with naturally occurring reflections and varied illumination, viewpoints, poses, eyeglasses, and reflection severity.They have no corresponding clean images and contribute to no-reference and blind-user evaluation.Source: §A.3.2 · p. 21

A.3.3 Real-world Test-set Characteristics

• Across the 458 reflected inputs, paired-reference and in-the-wild subsets have broad, distinct distributions in resolution, aspect ratio, luminance, contrast, high-frequency detail, and color saturation.The clean references from paired images are excluded from this image-statistics analysis.Source: §A.3.3 · pp. 22–23, Figure 9 · p. 22

B.1 Ocular-Prior-Guided Adaptive Representation

• OcuFlow builds an observation-only ocular importance field and structural supports from facial landmarks, then uses them to guide an adaptive quadtree while retaining native content and continuous leaf geometry.The importance weights prioritize the iris, followed by eyelid and spectacle-lens regions; sparse anchors preserve critical ocular locations.Source: §B.1.1 · pp. 23–24, §B.1 · p. 23

B.1.2 Exact-Budget Quadtree Allocation

• The exact-budget solver allocates 256 leaves by selecting 64 feasible splits with maximum importance mass, using a max-heap that remains optimal under structural constraints.Each split adds three leaves, and mandatory splits enforce maximum leaf sizes over active iris, eyelid, lens, or anchor supports.Source: §B.1.2 · pp. 25–26

B.2.1 Geometry and Cross-Scale Conditioning

• Geometry-aware cross-attention injects observation features into relevant tokens and lets fine ocular tokens draw on coarse facial context while leaving coarse tokens unchanged.Continuous leaf geometry informs attention queries and keys, while appearance features supply values.Source: §B.2.1 · p. 28

B.2.2 Observation-Anchored Source

• The observation-anchored source scales perturbation with non-background ocular occupancy while guaranteeing fine leaves a modulation weight of at least 0.75.A fixed σ = 0.05 keeps the perturbed source close to the reflected observation.Source: §B.2.2 · p. 29

B.2.3 pMF Parameterization

• pMF uses the observation-anchored source at t = 1 for one-evaluation restoration, while training samples both finite MeanFlow intervals and collapsed zero-length intervals.The default interval-sampling settings are μ = −0.4, σ_t = 1.0, and p_FM = 0.5.Source: §B.2.3 · p. 29, §B.2.4 · p. 29

B.3 Native-Resolution Frequency-Preserving Synthesis

• Native-resolution synthesis combines absolute and residual reconstructions in the frequency domain, with a residual gate adjusting the mix and lens-supported composition preserving the input outside the editable region.The spatial maps mix low-frequency absolute and residual reconstructions with high-frequency input and absolute reconstruction, respectively.Source: §B.3 · p. 30, Figure 11 · p. 30

B.3.1 Native Projection and Lens Support

• Native projection resamples canonical patches onto image supports and blends overlapping predictions with raised-cosine windows to suppress seams.A separate, feathered lens-support map avoids inheriting quadtree boundaries.Source: §B.3.1 · p. 30

B.3.2 Spatial Gate Construction

• Spatial gates distinguish regions needing reconstruction from those where native detail can be retained; correction magnitudes below 0.03 provide little reconstruction evidence, while those above 0.10 saturate the gate.The native-resolution low-pass operator is a 15 × 15 box filter.Source: §B.3.2 · pp. 30–31
• Analytical spatial gates favor clean reconstruction in eye, iris, and strong-highlight regions, while a lightweight learned gate refines only the two mixing maps.The analytical high-frequency gate initially preserves native detail with weights of 0.90 over ordinary lens regions, 0.30 over the eye, and 0.15 over the iris; the learned gate has 8,578 parameters and is zero-initialized to reproduce the analytical gates.Source: §B.3.2 · pp. 31–32

C.1.1 Training Setup and Objectives

• Training combines velocity-space restoration with image-space reconstruction, ocular-fidelity, and preservation constraints, plus a separate frequency-aware renderer objective.Training runs for 60k steps: 45k with the renderer frozen, 5k optimizing the renderer gate with restoration frozen, then 10k of joint fine-tuning.Source: §C.1.1 · pp. 32–34

C.1.2 Inference Details

• At inference, OcuFlow maps ocular geometry to a 256 × 256 domain, restores 256 canonical patches with one network evaluation, and renders them onto the padded observation.The renderer combines the restored patches with the padded input, after which the padded region is removed to recover the original image extent.Source: §C.1.2 · p. 34

C.1.3 Evaluation Protocols

• Because full-image PSNR can underweight eye-region errors, evaluation reports eyeglass-region and mixed-region metrics alongside ocular perceptual and landmark measures.Eye-ROI LPIPS measures perceptual similarity, while Eye-NME and Iris-NME quantify eye-contour and iris landmark displacement.Source: §C.1.3 · pp. 34–35
• Iris localization is evaluated with a shared, unfine-tuned detector; the real-world set uses NIQE and CLIP-IQA+, and a blind study gathers 3,990 selections.The study uses 30 images and 133 participants, with outputs shown in randomized order and method identities hidden.Source: §C.1.3 · pp. 36–37

C.1.4 Baseline Configurations

• The comparison includes Stable Diffusion 1.5 and 3.5, ER2Net, DSRNet, DSIT, RDNet, ReflexSplit, and DL2G; learned baselines use OcuBench training data.Stable Diffusion 1.5 uses a ReGlass-IP Lite adapter, while Stable Diffusion 3.5 uses task-adapted ControlNet.Source: §C.1.4 · p. 37

C.1.5 Efficiency Protocol

• Runtime comparisons use the same RTX 4090 and include each method’s complete inference pipeline, excluding file encoding and disk I/O.OcuFlow is timed at batch size 1, BF16, and 1-NFE; ocular-geometry extraction runs concurrently with GPU inference.Source: §C.1.5 · p. 38

C.2 Additional Quantitative Analysis

• OcuFlow ranks first on six of seven paired-reference metrics and second on Mixed-SSIM, while improving reconstruction, ocular fidelity, and geometry over competing methods.Against DSIT, it improves Mask-PSNR by 0.5654 dB and reduces Eye-ROI LPIPS by 29.2%; against RDNet, it reduces Eye-NME by 10.2%.Source: §C.2 · pp. 38–39
• At 56.4551 ms/image, OcuFlow is approximately 4.3× faster than ER2Net and 54.8× faster than DSIT, using 64.9% less GPU memory than DSIT.Source: §C.2 · p. 38
• NIQE is lowest at 5.2369 versus DSRNet’s 5.5638, CLIP-IQA+ ranks second at 0.6419, and OcuFlow wins 67.32% of 3,990 blind selections.The reflected input scores 0.6556 on CLIP-IQA+, illustrating that generic no-reference metrics do not directly measure reflection-removal correctness.Source: §C.2 · pp. 38–39, Table 10 · p. 39, Table 9 · p. 39

C.3 Downstream Iris Localization

• At IoU threshold 0.5, OcuFlow localizes both eyes successfully in 85.30% of images, versus 54.43% for reflected inputs and 83.78% for DSIT.The minimum left/right iris IoU has a median of 0.765 for OcuFlow, and the fixed detector is used without fine-tuning.Source: §C.3 · pp. 39–40, Figure 12 · p. 40

C.4 Additional Qualitative Results

• The additional paired examples span diverse subjects, eyeglass styles, reflection intensities, and ocular appearances, with enlarged crops for inspecting residual glare and detail preservation.Source: §C.4 · p. 40

C.4.1 Qualitative comparison.

• ER2Net and DSRNet often leave residual glare or reduced contrast, while DSIT and ReflexSplit can smooth ocular appearance and RDNet can retain chromatic residuals.Stable Diffusion outputs may alter local color, contrast, or facial appearance; DL2G can introduce inconsistent structures near rims and eyelids.Source: §C.4.1 · pp. 40, 42
• Across the shown cases, OcuFlow more closely follows clean references, removing localized glare while keeping eyeglass boundaries and surrounding facial texture stable.Under severe occlusion, iris position, eyelid contour, and local eye contrast remain closer to the clean reference than in several competing outputs.Source: §C.4.1 · p. 42

C.4.2 Ocular-prior-guided selective restoration.

• OcuFlow leaves non-overlapping lens reflections largely unchanged but applies stronger correction when a bright reflection obscures an eye.The ocular prior concentrates restoration on reflections overlapping visible ocular structures, reducing unnecessary changes elsewhere on the lenses.Source: §C.4.2 · pp. 42–43, Figure 16 · p. 43

C.4.3 Failure cases.

• Dark sunglasses can trigger reconstruction of visible eye content behind attenuating lenses, producing results that diverge from the clean reference.Source: §C.4.3 · p. 44, Figure 17 · p. 44

C.5 Additional Ablation and Inference Analysis

• The additional ablations examine training-data sampling, auxiliary objectives, semantic conditioning, renderer gates, and multi-step inference.Source: §C.5 · p. 44

C.5.1 Training and Conditioning Ablations

• Natural sampling outperforms balanced source sampling across all seven reported metrics, favoring the empirical training-data mixture.Source: §C.5.1 · p. 44
• Removing the eye-gradient objective, semantic prior, or preservation objective produces smaller but consistent degradation across reconstruction, perceptual, and ocular-geometry metrics.Source: §C.5.1 · p. 44
• Removing the multi-scale reconstruction term lowers Mask-PSNR from 22.1054 to 21.3276 dB and raises Eye-NME from 1.4030% to 1.6258%.Eye-ROI LPIPS decreases from 0.1035 to 0.1015, showing that this single perceptual metric diverges from the pixel-fidelity and ocular-geometry changes.Source: §C.5.1 · pp. 44–45

C.5.2 Frequency-Renderer Gate Ablation

• Learned residual gates improve six of eight reported metrics, including Mask-PSNR by 0.2079 dB and Out-PSNR from 26.2977 to 26.4147 dB.Out-PSNR is computed outside the eyeglass evaluation mask.Source: §C.5.2 · p. 45
• The fixed gates yield slightly lower Eye-ROI LPIPS and Iris-NME, while learned gates provide stronger masked and mixed reconstruction fidelity with comparable ocular quality.Source: §C.5.2 · p. 45

C.5.3 Multi-step Inference

• Additional MeanFlow evaluations do not progressively improve reference-aligned metrics; 1-NFE gives the highest Mask/Mixed PSNR and SSIM and lowest ocular NME.Runtime rises from 56.4551 ms/image at 1-NFE to 72.6149 ms/image at 8-NFE, while multi-step inference shifts the trade-off toward lower LPIPS.Source: §C.5.3 · pp. 45–46, Table 13 · p. 46

C.5.4 Inference-time Breakdown and Pipeline Parallelism

• The predictor accounts for 12.5873 ms (22.3%) and the renderer 14.1714 ms (25.1%), while CPU preparation is the largest individual contribution at 17.6631 ms (31.3%).Within rendering, partition-of-unity projection accounts for 66.1% of rendering cost, versus 33.6% for frequency/decode operations and 0.3% for lens resizing.Source: Table 14 · p. 46, §C.5.4 · p. 46
• Four asynchronous preprocessing workers yield an idealized preparation interval of approximately 55.0956 ms/image from 220.3822 ms/image of aggregate CPU work.Without CPU–GPU overlap, the preparation and inference stages yield an estimated 93.8876 ms/image interval.Source: §C.5.4 · p. 46
• 56.4551 ms/image is within approximately 2.5% of the 55.0956 ms/image idealized full-overlap interval under sustained pipelined execution.The idealized interval is set by the slower of the pipeline stages.Source: §C.5.4 · p. 46