Research map / Video / Video segmentation and tracking
Making Old Film Great Again: Degradation-aware State Space Model for Old Film 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 MambaOFR, a recurrent Mamba-based old film restoration model with degradation-aware prompts and flow-guided mask deformable alignment, plus a new synthetic/real-world old film benchmark (SRWOV).
Problem
Old film restoration must handle complex, composite analog and digital degradations plus structured defects (scratches, cracks, stripes) that propagate across frames; existing specialized methods lag behind general video restoration methods, and no unified public benchmark exists.
Approach
Combines a 2-D selective state space (Mamba) recurrent restoration network with a degradation-aware prompt block that generates dynamic convolution parameters from global context, and a flow-guided mask deformable alignment module that predicts soft masks to stop structured defects from propagating through time.
Key results
- On the synthetic SRWOV part, MambaOFR reaches 26.93 dB PSNR and 0.9271 SSIM, the best among compared methods (RTN 25.10 dB, RVRT 24.92 dB).
- MambaOFR reports LPIPS 0.0538 and DISTS 0.0583 on synthetic data, improving perceptual metrics over the best competitor by about 46% per the authors.
- On the real-world part, MambaOFR achieves the best NIQE (4.8261) and BRISQUE (19.1769) and a CLIPIQA+ of 0.4142.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MambaOFR instead of ShiftNet | Old film restoration · SRWOV synthetic | PSNR (dB): 26.0208 -> 26.9340 | Helped |
| use MambaOFR instead of RTN | Old film restoration · SRWOV synthetic | LPIPS: 0.1025 -> 0.0538 | Helped |
| use MambaOFR instead of RTN | Old film restoration · SRWOV real-world | BRISQUE: 28.9296 -> 19.1769 | Helped |
| add the Mamba block instead of a Swin-Transformer block (ablation) | MambaOFR · SRWOV synthetic | PSNR (dB): 26.1231 -> 26.9340 | Helped |
| add the degradation-aware prompt block instead of standard convolution and attention (ablation) | MambaOFR · SRWOV synthetic | PSNR (dB): 26.1432 -> 26.9340 | Helped |
| add flow-guided mask deformable alignment instead of mask-based alignment (ablation) | MambaOFR · SRWOV synthetic | PSNR (dB): 26.3467 -> 26.9340 | Helped |
| increase prompt parameter groups from 4 to N=16 (ablation) | MambaOFR · SRWOV synthetic | PSNR (dB): 26.3546 -> 26.9340 | Helped |
| increase prompt parameter groups from 8 to N=16 (ablation) | MambaOFR · SRWOV synthetic | PSNR (dB): 26.5891 -> 26.9340 | Helped |
Surprising findings
- General video restoration methods achieve higher PSNR/SSIM than specialized old film methods, while specialized methods do better on perceptual metrics LPIPS and DISTS.
Limitations
- Struggles to align features in heavily polluted frames where adjacent-frame motion estimation fails.
- Recovering fine details in highly degraded videos remains challenging, giving suboptimal results.
Experimental setup
Datasets: REDS SRWOV (proposed)
Benchmarks: SRWOV synthetic part SRWOV real-world part
Metrics: PSNR SSIM LPIPS DISTS NIQE BRISQUE CLIPIQA+
Baselines: DeOldify OldPhoto DeepRemaster RTN VRT RVRT ShiftNet
Builds on: Mamba vision mamba dynamic convolution RTN RAFT-style optical flow
Closest papers by meaning
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.85)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.85)
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.84)
- Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.84)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.84)
- Decoupling Degradations with Recurrent Network for Video Restoration in Under-Display Camera (AAAI 2024 · similarity 0.84)
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.