Research map / Image generation / Image restoration
WaveFace: Authentic Face Restoration with Efficient Frequency Recovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
WaveFace performs blind face restoration in the wavelet frequency domain, using a conditional diffusion model on the low-frequency sub-band and a one-pass U-shaped network for high-frequency sub-bands, achieving ~10x faster inference than pixel-domain diffusion BFR methods.
Problem
Diffusion-based blind face restoration is slow to train and infer on 512x512 images and fails to preserve identity and fine facial details because the unconditional generation is uncertain.
Approach
Decompose faces with Haar DWT (level 2): run a diffusion model conditioned on the LQ low-frequency sub-band (concatenated at each denoising step) on a 128x128 low-frequency map, and recover all high-frequency sub-bands at multiple levels in a single forward pass with a U-shaped network, then reconstruct via IWT.
Key results
- On CelebA-Test WaveFace reaches PSNR 26.967 dB, SSIM 0.711, LPIPS 0.343, and identity angle 28.263 degrees, the best identity preservation (6.49 degrees better than the second-best method).
- Inference time drops to 1.97 s at DWT level 2 (128x128 DM input) versus 25.04 s for DifFace, about 10x faster than existing diffusion BFR methods.
- On real-world sets WaveFace achieves the best FID on LFW (43.175) and WIDER (36.913); on WebPhoto (81.525) it beats DifFace-style diffusion baselines but not the best GAN-based method (GPEN 80.572 is lower).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DWT level 1 instead of level 2 (ablation) | WaveFace · CelebA-Test | PSNR: 26.97 -> 27.08 | Helped |
| Use DWT level 3 instead of level 2 (ablation) | WaveFace · CelebA-Test | PSNR: 26.97 -> 24.81 | Hurt |
| Increase DWT level from 2 to 3 (ablation) | WaveFace · CelebA-Test | inference time: 1.97 -> 1.46 | Helped |
| Add the LQ low-frequency image as a condition by concatenation (ablation) | WaveFace LCD · CelebA-Test | PSNR: 9.77 -> 26.97 | Helped |
| Add HFR at both DWT levels (ablation) | WaveFace · CelebA-Test | PSNR: 24.353 -> 26.967 | Helped |
| Add HFR at DWT level 2 (ablation) | WaveFace · CelebA-Test | PSNR: 24.819 -> 26.967 | Helped |
| Add HFR at DWT level 1 (ablation) | WaveFace · CelebA-Test | PSNR: 26.081 -> 26.967 | Helped |
| Add HFR at both DWT levels (ablation) | WaveFace · CelebA-Test | SSIM: 0.430 -> 0.711 | Helped |
| Replace concatenation conditioning with AdaGN identity embedding (ablation) | WaveFace LCD · CelebA-Test | PSNR: 26.97 -> 21.55 | Hurt |
| Replace concatenation conditioning with cross-attention (ablation) | WaveFace LCD · CelebA-Test | PSNR: 26.97 -> 25.85 | Hurt |
| Use WaveFace instead of RestoreFormer | blind face restoration · CelebA-Test | PSNR: 24.191 -> 26.967 | Helped |
| Use WaveFace instead of DifFace | blind face restoration · LFW-Test | FID: 45.227 -> 43.175 | Helped |
Surprising findings
- Removing the LQ condition and using an unconditional diffusion model drops identity preservation and image quality sharply (PSNR 9.77 dB), and the authors attribute this to slow convergence.
- Going to DWT level 3 hurts restoration quality (PSNR 24.81) even though it is fastest, showing a sharp efficiency-quality trade-off.
Limitations
- Gap between real-world degradations and the simulated degradation model leads to inferior results in some cases.
- Mosaic artifacts and information loss appear at higher DWT levels (J>=3).
Experimental setup
Datasets: FFHQ CelebA-HQ
Benchmarks: CelebA-Test LFW-Test WebPhoto-Test WIDER-Test
Metrics: PSNR SSIM LPIPS FID identity angle (ArcFace Deg.) parameters inference time
Baselines: GPEN GFP-GAN VQFR RestoreFormer CodeFormer DR2 DifFace SwinIR
Builds on: DDPM DDIM DifFace Haar wavelet transform
Closest papers by meaning
- FaceMe: Robust Blind Face Restoration with Personal Identification (AAAI 2025 · similarity 0.89)
- OSDFace: One-Step Diffusion Model for Face Restoration (CVPR 2025 · similarity 0.87)
- ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration (NeurIPS 2024 · similarity 0.87)
- Overcoming False Illusions in Real-World Face Restoration with Multi-Modal Guided Diffusion Model (ICLR 2025 · similarity 0.86)
- Feature out! Let Raw Image as Your Condition for Blind Face Restoration (ICML 2025 · similarity 0.86)
- Self-Supervised Selective-Guided Diffusion Model for Old-Photo Face Restoration (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.