- 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
- 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.
- 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.
- Efficiency potential. Unlike diffusion-based competitors, GeoFAME restores in a single forward pass — a practical advantage if substantiated.
Critical weaknesses
- 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.
- 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.
- 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.
- 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?