Research map / Image generation / Image restoration
Feature out! Let Raw Image as Your Condition for Blind Face Restoration
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.
Proposes P-I2SB, a Schrödinger bridge framework for blind face restoration that pseudo-hashes HQ training images using LQ counterparts so the raw LQ image can serve directly as the diffusion endpoint.
Problem
Blind face restoration maps low-quality face images with unknown, complex degradations to high-quality images; existing diffusion methods rely on feature extractors to guide from LQ inputs, and vanilla Image-to-image Schrödinger Bridge (I2SB) fails because one HQ image maps to many LQ images, violat
Approach
Preprocess HQ training images into pseudo-samples via a rule tied to their paired LQ image (concatenation, residual, or light noise), so the endpoint distributions become structurally consistent and an SB between pseudo-HQ and raw LQ images is well posed and reversible; inference inverts the hashing.
Key results
- On CelebA-Test, P-I2SB (Cat-I2SB variant) reports FID 13.910 and LPIPS 0.2395, the best among compared methods, with PSNR 25.405 and SSIM 0.6581.
- Ablation on CelebA-Test: Cat hashing reduces NIQE from 25.49 (no hashing, condition only) to 13.91 and FID from 6.39 to 5.33.
- On real-world Wider, P-I2SB achieves NIQE 4.9056 and FID 69.62, leading the compared methods on these metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LQ images as guidance (ablation) | Vanilla-I2SB · CelebA-Test | NIQE: 6.5709 -> 6.4849 | Helped |
| add LQ images as guidance (ablation) | Vanilla-I2SB · CelebA-Test | LPIPS: 0.2851 -> 0.2904 | Hurt |
| add degradation-aware representation alongside LQ guidance (ablation) | Vanilla-I2SB · CelebA-Test | FID: 25.6026 -> 25.4932 | Helped |
| use Noise pseudo-hashing (ablation) | I2SB with LQ and degradation-aware guidance · CelebA-Test | FID: 25.4932 -> 18.2646 | Helped |
| use Cat pseudo-hashing (ablation) | I2SB with LQ and degradation-aware guidance · CelebA-Test | FID: 25.4932 -> 13.9109 | Helped |
| use Res pseudo-hashing (ablation) | I2SB with LQ and degradation-aware guidance · CelebA-Test | FID: 25.4932 -> 14.2941 | Helped |
| use Cat rather than Noise pseudo-hashing (ablation) | P-I2SB · CelebA-Test | FID: 18.2646 -> 13.9109 | Helped |
| use Res rather than Cat pseudo-hashing (ablation) | P-I2SB · CelebA-Test | FID: 13.9109 -> 14.2941 | Hurt |
Surprising findings
- Vanilla I2SB performs worse than many guidance-based baselines on real-world sets (e.g., Wider FID 108.03 vs 69.62 for P-I2SB), confirming the endpoint issue in blind settings.
Limitations
- Synthetic degradation pipeline does not fully replicate real-world degradations such as watermarks, where restoration remains suboptimal.
- Training data dominated by frontal faces limits restoration of extreme poses.
Experimental setup
Datasets: FFHQ CelebA-HQ CelebA-Test Real-World LFW CelebChild WebPhoto-Test Wider
Benchmarks: CelebA-Test LFW CelebChild WebPhoto-Test Wider
Metrics: SSIM PSNR FID NIQE LPIPS
Baselines: PSFRGAN GFPGAN GPEN VQFR CodeFormer RestoreFormer DMDNet DAEFR DifFace DR2 PGDiff DiffBIR
Builds on: I2SB DDPM DDIM
Closest papers by meaning
- ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration (NeurIPS 2024 · similarity 0.86)
- WaveFace: Authentic Face Restoration with Efficient Frequency Recovery (CVPR 2024 · similarity 0.86)
- AsyncDSB: Schedule-Asynchronous Diffusion Schrödinger Bridge for Image Inpainting (AAAI 2025 · similarity 0.86)
- Finding Local Diffusion Schrodinger Bridge using Kolmogorov-Arnold Network (CVPR 2025 · similarity 0.86)
- Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge (ICML 2024 · similarity 0.85)
- Unpaired Image-to-Image Translation via Neural Schrödinger Bridge (ICLR 2024 · 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.