Research map / Image generation / Image restoration
Self-Supervised Selective-Guided Diffusion Model for Old-Photo 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.
Training-free old-photo face restoration that guides a frozen FFHQ-pretrained face diffusion model with self-supervised pseudo-references generated under weak guidance, applying restoration guidance throughout and color guidance in later steps.
Problem
Restoring old photos of faces with compounded degradations (breakage, fading, severe blur), where learned BFR models and existing zero-shot diffusion guidance methods fail on unseen breakage and face color.
Approach
Uses pseudo-reference faces produced by a pre-trained face diffusion model under weak guidance as self-supervision, with staged, region-selective losses (face parsing and scratch masks) that guide breakage completion throughout denoising and skin color refinement in later steps.
Key results
- On VintageFace medium split, SSDiff reports FID 128.31 vs 136.73 for PGDiff and 132.14 for DifFace, with TOPIQ 0.6412 (best).
- On the hard split, SSDiff reaches FID 122.59 and BRISQUE 8.95, best among compared methods.
- Identity distance to LQ inputs is 0.2280 for SSDiff, second best behind DiffBIR (0.2223) and better than PGDiff (0.2859).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add self-supervised pseudo-label guidance (ablation) | SSDiff · VintageFace medium | FID: 145.7 -> 128.3 | Helped |
| Add self-supervised pseudo-label guidance (ablation) | SSDiff · VintageFace medium | TOPIQ: 0.618 -> 0.641 | Helped |
| Add self-supervised pseudo-label guidance (ablation) | SSDiff · VintageFace medium | BRISQUE: 7.92 -> 7.29 | Helped |
| Add L1 intact-region fidelity guidance (ablation) | SSDiff · VintageFace medium | FID: 142.8 -> 128.3 | Helped |
| Add L1 intact-region fidelity guidance (ablation) | SSDiff · VintageFace medium | TOPIQ: 0.602 -> 0.641 | Helped |
| Add L1 intact-region fidelity guidance (ablation) | SSDiff · VintageFace medium | BRISQUE: 8.18 -> 7.29 | Helped |
| Add L2 breakage smoothness guidance (ablation) | SSDiff · VintageFace medium | FID: 132.1 -> 128.3 | Helped |
| Add L3 selective coloring guidance (ablation) | SSDiff · VintageFace medium | FID: 136.3 -> 128.3 | Helped |
| Add L3 selective coloring guidance (ablation) | SSDiff · VintageFace medium | TOPIQ: 0.624 -> 0.641 | Helped |
| Add L3 selective coloring guidance (ablation) | SSDiff · VintageFace medium | BRISQUE: 7.65 -> 7.29 | Helped |
| Use SSDiff instead of CodeFormer | old-photo face restoration · VintageFace simple | BRISQUE: 6.69 -> 6.71 | No clear effect |
| Use SSDiff instead of DifFace | old-photo face restoration · VintageFace hard | FID: 145.85 -> 122.59 | Helped |
Surprising findings
- Code (CodeFormer) achieves the best BRISQUE on the simple split (6.69) over SSDiff (6.71).
- Weak guidance pseudo-references deviate from original identity yet provide natural color and breakage completion, which the authors leverage as supervision.
Limitations
- Large stained regions may be misinterpreted as facial or hair structure and restored incorrectly.
- Evaluation relies on no-reference metrics since ground truth is unavailable; identity fidelity computed only on 50 selected inputs.
Experimental setup
Datasets: FFHQ
Benchmarks: VintageFace
Metrics: FID BRISQUE TOPIQ MAN-IQA CLIP identity distance
Baselines: BOPB CodeFormer DifFace DiffBIR DDNM PGDiff
Builds on: DDPM classifier guidance PGDiff DDNM
Closest papers by meaning
- ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration (NeurIPS 2024 · similarity 0.88)
- Overcoming False Illusions in Real-World Face Restoration with Multi-Modal Guided Diffusion Model (ICLR 2025 · similarity 0.88)
- OSDFace: One-Step Diffusion Model for Face Restoration (CVPR 2025 · similarity 0.87)
- PFStorer: Personalized Face Restoration and Super-Resolution (CVPR 2024 · similarity 0.87)
- $InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration (ICLR 2025 · similarity 0.86)
- SHMT: Self-supervised Hierarchical Makeup Transfer via Latent Diffusion Models (NeurIPS 2024 · similarity 0.86)
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.