An arXiv preprint reports a physics-grounded computer-vision pipeline designed to remove eyeglasses from images and video while targeting preservation of facial identity, expression, pose and temporal consistency. On the clear-glasses portion of a manually curated FFHQ evaluation, the proposed method reported an FID of 0.379, pupil displacement of 2.249 ± 1.957 pixels, landmark L2 error of 0.632 ± 0.271 pixels and an inference speed of 27.68 frames per second.
Those figures are results from a selected computer-vision evaluation, not evidence that the method works reliably on every real-world face, video or prescription lens. The paper is an arXiv version 1 preprint dated 20 Aug 2026.
A synthetic path to a glasses-free face
The training data were synthetic rather than a set of naturally paired recordings. The curated dataset contained 1,860 identities and 24,180 image pairs. Its data-creation workflow used a three-stage hard-sampling filter, including background and ocular L1 filtering for shifts in head pose, gaze and eye identity.
During training, the pipeline simulated optical lens effects on curated generative face sets. The purpose was to create paired inputs that were both consistent and diverse, giving the system examples of faces before and after the simulated glasses effects.
The model, called JFSnet, maps an input image to a clean RGB face through a pre-trained DINOv2 ViT-L/14 encoder and a ResNet-based convolutional decoder. The architecture therefore combines a transformer-based image encoder with a convolutional reconstruction stage, as described in the paper.
The training also enforced translation equivariance, a constraint intended to encourage the output to remain temporally stable when image content shifts position. That design choice became the focus of a separate ablation in the supplementary results.
What the benchmark numbers show
The still-image evaluation used two manually curated FFHQ groups: 12,163 clear-glasses images and 2,176 sunglasses images. For the clear-glasses group, the reported scorecard included FID, pupil displacement, landmark L2 error and inference speed. The proposed method’s values were 0.379, 2.249 ± 1.957 pixels, 0.632 ± 0.271 pixels and 27.68 frames per second, respectively.
FID is a distribution-level image metric, so it can summarise how two sets of images compare but does not directly establish whether a person’s identity or facial geometry has been preserved. The supplied analysis also does not define what the reported ± values represent, leaving their precise interpretation unclear.
For video, the authors used a CelebV-Text temporal protocol. JFSnet reported RAFT-L1 of 7.682, temporal LPIPS of 18.528 and Laplacian variance of 162.6. The paper presents these as temporal-stability and image-quality indicators, but no uncertainty estimates are reported for the three values.
In the translation-equivariance ablation, RAFT-L1 changed from 8.023 to 7.682 and tLPIPS changed from 19.520 to 18.528. The table reports no degradation in spatial restoration alongside those changes. Because this was an architectural ablation rather than a randomized experiment, the comparison does not establish that the component caused the observed differences in other settings.
A separate architectural ablation reported FID of 0.379 for the proposed ViT-CNN configuration. That figure records the result for the configuration tested in the supplementary comparison; it does not by itself establish universal superiority over other designs.
A small preference test favored the proposed output in several comparisons
The study included a perceptual test using 60 CelebV-Text video clips, 60 FFHQ portraits and 37 participants. Four other comparison methods were included, with five methods shown per question.
In the Nano Banana comparison row of the preference heatmap, the proposed method received 97%, 100% and 100% across the listed video criteria. Across the five listed image criteria, the reported figures were 81%, 42%, 75%, 55% and 74%.
The percentages are criterion-specific results from that survey, not a single overall accuracy figure. The supplied analysis reports no confidence intervals, participant-level variability or formal significance tests, so the results do not show that the preference pattern would hold beyond the selected examples and the 37-person study.
The real-world test is still missing
The main limitation is the distance between the training setup and naturally occurring footage. The method relies on synthetic face pairs and simulated lens optics, while the FFHQ evaluation uses manually curated image subsets. The study therefore does not establish performance on real paired videos with glasses-free ground-truth frames or on arbitrary real-world prescription lenses.
The paper reports four categories of failure: incomplete eyewear removal, residual shadows from frames, minor iris deformation and undercorrection of refraction. The supplied analysis also notes simplifying assumptions in the optical simulation, including monocular depth, standard head scale, portrait-camera geometry and simplified lens geometry.
The benchmark does not answer every identity question either. FID does not directly measure identity preservation or geometric correctness, and the study’s reported survey preferences do not establish biometric identity preservation. The comparisons were also limited by non-exhaustive baseline tuning, while Runway Gen-4.5 appeared only in the preference study.
The authors’ results are best read as evidence that the proposed pipeline can be evaluated on selected synthetic, image and video protocols, with reported signs of temporal and perceptual performance. They do not establish causal superiority over the comparison methods or guarantee that the same results will appear on broader data.
What comes next
The open tests are practical ones: real paired videos with glasses-free reference frames, unseen high-power prescriptions, non-spherical lenses, profile views and unusual lighting. Independent evaluation of the reported failure cases would also show whether the method’s weaknesses are occasional edge cases or recurring problems.
Reproducibility is another unresolved issue because the workflow depends on a synthetic-data process and simulated optics. The supplied record states that a project site is available, but it does not report whether the synthetic dataset, code or trained weights have been released.
The paper reports partial funding from BMK, BMAW, the State of Upper Austria through SCCH INTEGRATE within the FFG COMET programme, and Czech Technical University in Prague grant SGS26/074/OHK3/1T/13.
Paper data and sources
Original title: Unwarping the Lens: A Physics-Grounded Approach to Video Glasses Removal
Authors: Radim Spetlik, David Futschik, Radek Danecek et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text