HiGlassRM: Learning to Remove
High-prescription Glasses
via Synthetic Dataset Generation

Sebin Lee and Heewon Kim
Soongsil University
WACV 2026
Paper Video Data (Soon)
FFHQ Dataset Examples Our HiGlassRM explicitly compensates this geometric distortion,
preserving identity-consistent facial geometry and background alignment.


Abstract

Existing eyeglass removal methods can handle frames and shadows but fail to correct lens-induced geometric distortions, as public datasets lack the necessary supervision. To address this, we introduce the HiGlass Dataset, the first large-scale synthetic dataset providing explicit flow-based supervision for refractive warping. We also propose HiGlassRM, a novel pipeline whose core is a network that explicitly estimates a displacement flowmap to de-warp distorted facial geometry. Experiments on both synthetic and real images show that this flowmap-centric approach, trained on our data, significantly improves identity preservation and perceptual quality over existing methods. Our work demonstrates that explicitly modeling and correcting geometric distortion via flowmap estimation, enabled by targeted supervision, is key to faithful eyeglass removal.

Synthetic Dataset Generation

Synthetic Dataset Generation Figure3: Overview of the HiGlass Dataset synthesis.

  • 1️⃣Segment lens area from the binary eyeglass-frame mask ($M$)
  • 2️⃣Depth-& center-aware flowmap synthesis
  • $$F(\mathbf{p}) = (\mathbf{p} - \mathbf{p}_c) \odot R_{\text{depth}}(\mathbf{p}) \odot S(\mathbf{p})$$
  • 3️⃣Selective warping ($𝐿_𝑑$)+ background-preserving composition


Synthetic Dataset Generation Results Figure6: Visual examples from HiGlass Dataset.

  • The HiGlass Dataset contains 29,071 paired samples, each consisting of five components ($I$, $M$, $F$, $D$, $O$).

HiGlassRM Framework

Method Overview Figure4: Overview of the proposed HiGlassRM.

  • HiGlassRM decouples geometric correction from appearance restoration.
  • The core component, the Flowmap Network, predicts a two-channel displacement field $\widehat{F}$ to explicitly model lens-induced distortion.
  • The predicted flow is applied via grid sampling to restore the warped facial geometry.
  • Once geometry is restored, the De-Glass Network synthesizes the final eyeglass-free output.

Experimental Results on Real Data

Dataset Examples Figure 8. Qualitative comparison on the MeGlass dataset

Experimental Results on HiGlass Dataset

Dataset Examples Figure 7. Qualitative comparison of our method (HiGlassRM) against five baseline methods and the Ground Truth on the HiGlass Dataset. Dataset Examples Table 1. Quantitative comparison on the test split of the HiGlass dataset.

Citation

TBD (To be published at WACV 2026)

Acknowledgements

TBD (To be published at WACV 2026)