acceptodds
Withdrawn from ICLR 2027

GeoFAME: Restoring Beautified Faces via Geometry Correction and Frequency-Aware Multi-Expert Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

With the widespread use of social media platforms and intelligent portrait editing tools, facial beautification has become a common form of image manipulation, raising potential concerns about the authenticity of face images and identity consistency. Existing studies mainly focus on detecting beautified face images, while how to accurately recover their corresponding natural faces remains underexplored. Operations such as face slimming, eye enlargement, skin smoothing, and skin whitening simultaneously alter facial geometry, local texture, and color distribution, making facial beautification restoration a challenging task. To address this problem, we propose GeoFAME, a Geometry-Corrected Frequency-Aware Multi-Expert Network for facial beautification restoration. GeoFAME first alleviates non-rigid deformations introduced by beautification through geometric correction, and then employs a Laplacian pyramid to decompose multi-scale frequency information, with structure, texture, and color experts performing targeted restoration. Meanwhile, a spatially adaptive router dynamically fuses the expert outputs according to the restoration requirements of different facial regions. Extensive experiments demonstrate that GeoFAME effectively restores structural, textural, and color changes caused by both single and composite beautification operations, while achieving favorable restoration performance and cross-dataset generalization under different beautification intensities.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Computer Vision, Facial Beauty Retouching Restoration, Multi‑Expert Network, Frequency‑Guided, Spatial‑Adaptive Routing, Laplacian Pyramid, Image Restoration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_xrwNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoFAME, a single-pass, non-diffusion framework for facial beautification restoration. It first corrects non-rigid deformations (face slimming, eye enlargement) via a predicted displacement field, then decomposes the corrected image with a four-level Laplacian pyramid into frequency bands processed by three residual experts (texture, structure, color). A spatially adaptive router fuses the expert residuals pixel-wise on top of a U-Net base restoration. GeoFAME reports the best performance on a self-collected Meitu-CelebA benchmark at 30% and 60% beautification strengths.

Critical strengths
  1. Decoupled-degradation design. Geometry, texture, and color artifacts are handled by dedicated branches and fused via learned pixel-wise weights. The frequency-to-expert assignment is interpretable, and the ablation (Table 4) confirms the three experts are complementary.
  2. Strong results on the proposed benchmark. Best PSNR/SSIM/FID at both strengths (Table 1), best single-operation cross-dataset results (Table 2), and the highest user-study scores on all three criteria (Table 3), against nine baselines including MoFRR and DiffClean.
  3. Efficiency potential. Unlike diffusion-based competitors, GeoFAME restores in a single forward pass — a practical advantage if substantiated.
Critical weaknesses
  1. No objective identity-consistency evaluation. Identity preservation is the core motivation, yet the only identity evidence is the subjective user-study score. Standard metrics (e.g., ArcFace cosine similarity to the reference, verification accuracy after restoration) are missing.
  2. No efficiency measurements. The claimed single-pass advantage over diffusion baselines is unsupported: parameters, FLOPs, inference time, and memory are not reported for any method.
  3. Limited generalization. Training uses only 1,500 Meitu-retouched CelebA images from a single beautification tool. On the cross-dataset Multi setting (Table 2), GeoFAME trails DiffClean on SSIM (0.8684 vs. 0.9628) and LPIPS (0.0861 vs. 0.0335), while the text highlights only PSNR/FID gains.
  4. Ablation inconsistencies. The geometry-branch ablation is a separately retrained control whose numbers don't match the main model, and adding the branch worsens LPIPS/FID at 30%; no landmark error or displacement-field visualization verifies the deformation is corrected.
Questions for authors

The proposed method's PSNR increases from 32.15 dB (30%) to 35.67 dB (60%) — opposite to all baselines and counterintuitive, since stronger manipulation should be harder to invert. How were the two strength levels constructed, and what explains this?

Reviewer_q6jKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed Geometry-Corrected Frequency-Aware Multi-Expert Network (GeoFAME) for facial beautification restoration.

Critical strengths
  • The proposed method quantitatively outperforms the baselines on the RetouchFFHQ benchmark by MoE architecture with Texture, Structural, and Color Experts.
Critical weaknesses
  • Please provide a detailed description of the objective functions used to optimize the model. In the current manuscript, it is difficult to clearly understand what model in trained to learn, and what specific representations each expert is encouraged to learn.
  • PSNR measures pixel-level differences between predictions and ground truth (GT), while SSIM evaluates structural similiarity. To overcome the limitations of these metrics, perceptual metrics such as LPIPS and FID have been introduced. Although these metrics can assess similarity to beautified GT images, I believe there remains a gap between these metrices and the quantitative evaluation of perceptual beautification restoring quality. In this paper cases, although the proposed method achieves approximately 2.7 dB higher PSNR than MoFRR (ICCV 2025), the differences in LPIPS and FID are marginal. DiffClean (CVPR 2026) provides another example supporting this corcern. Despite achieving a higher PSNR than MoFRR, DiffClean exhibits substantially worse FID scores compared to both MoFRR and the proposed method.
  • To address the limitations of these metrics, recent studies such as [1] have adopted alternative evaluation metrics, such as ArcFace. Furthermore, some studies [2] define beautification in terms of specific improvements, such as acne removal and skin texture enhancement, rather than simply treating it as general skin retouching. Therefore, I believe the proposed method should be evaluated on a wider range of benchmarks to provide a more comprehensive assesment of its effectiveness.

[1] Jiachen Yang et al., BeautyGRPO: Aesthetic Alignment for Face Retouching via Dynamic Path Guidance and Fine-Grained Preference Modeling, CVPR 2025

[2] Zhidan Xu et al., Face Retouching with Diffusion Data Generation and Spectral Restorement, ICCV 2025

Questions for authors
  • I am concerned that, despite the introduction of Texture Loss and multi-scale frequency decomposition, the proposed method may still fail to faithfully reconstruct high-frequency details, such as freckles, in the original images, particularly given its high PSNR values. Could the authors provide representative examples that clearly demonstrate the differences between the baselines and the proposed method in reconstructing high-frequency facial details, such as freckles, acne, and deep pitted scars? It is well known that conventional pixel-wise losses, particualry MSE loss, can improve PSNR while generating overly smoothed or blurry images.
  • Except for PSNR, the quantitative performance differences between GeoFAME and prior methods, such as MoFRR and DiffClean, are marginal (Table 2), and the user study results (Table 3) further support this observation. I also find it difficult perceptually distinguish the results of MoFRR, DiffClean, and the proposed method in Figure 3. Therefore, I encourage the authors to clearify the key strengths and advantages of GeoFAME over prior work.
Reviewer_MXBMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets facial beautification restoration. The goal is to recover the natural face from a beautified one (whitened, smoothed, slimmed, eye-enlarged). GeoFAME first runs a geometric-correction branch. It predicts a displacement field and warps the input to undo non-rigid deformations. The corrected image is then handled two ways. (a) A four-level Laplacian pyramid splits it into high / mid-high / mid residuals and a low-frequency base. (b) A shared encoder produces features that drive a spatially adaptive softmax router, giving per-pixel expert weights. The weighted expert residuals are added back to the geometrically corrected image.

Critical strengths
  • Adding an explicit geometric-correction stage before appearance restoration is an intuitive decomposition.
  • The method does report competitive numbers on the authors' own benchmark, plus a user study (20 raters scoring identity / restoration / naturalness) that gives some subjective support.
Critical weaknesses
  • Missing methodology and unreliable ablations. No loss functions specified. The central claim is that the three experts are "complementary", which is not sound. The work is not easy to be reproduced purely through downstream reconstruction loss.
  • Weak and likely unfair comparison. The self-built dataset is small (1000/500 images). Several baselines are seemingly not retrained. Geometric correction branch is another core element. However, the geometric modifications are too easy to measure the effectiveness of geometric correction branch
Questions for authors
  • What exactly is the training objective (all loss terms and weights)? How is the displacement field supervised — are ground-truth correspondences available, or is it learned purely through downstream reconstruction?
  • Were all baselines retrained on the identical paired beautification-restoration data and protocol? If not, how is the comparison justified (especially MAD/SSAT/SCGAN, which are not restoration methods)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.