MgfrOFR

Official implementation of Multi-Granularity Feature Decomposition with Ranking-Aware Regularization for Old Film Restoration.

MgfrOFR is an old-film restoration framework built on a recurrent video restoration backbone. It introduces multi-granularity feature decomposition for degradation-aware restoration and ranking-aware temporal regularization for more stable video recovery.

Paper and pretrained weight links will be updated after the final camera-ready/release assets are frozen. The source code, evaluation scripts, and configuration files are already available in this repository.

News

• 2026-05-10: Initial public-release codebase prepared with training/testing configs and BasicOFR/BasicSR runtime.

Highlights

• Old-film restoration focus: designed for mixed analog degradations such as scratches, noise, blur, flicker, and local structured artifacts.
• Multi-granularity feature decomposition: adds MGFE modules to separate restoration cues at different feature granularities.
• Ranking-aware temporal regularization: provides a training loss for temporally consistent restoration behavior.
• Reproducible configs: includes full-model and ablation configs for training and testing.
• Standalone release: packages the required mgfrofr, basicofr, and basicsr code paths in one repository.

Method Overview

MgfrOFR architecture. The model extends a recurrent Mamba-based old-film restoration backbone with multi-granularity feature extraction and training-time ranking-aware regularization.

MGFE. Multi-granularity restoration-token branches summarize coarse-to-fine degradation cues and recalibrate propagated features.

RATR. Ranking-aware triplet regularization discourages branch collapse and promotes complementary temporal descriptors during training.

Qualitative Results

Synthetic REDS-style old-film restoration. MgfrOFR improves structural recovery and suppresses local old-film artifacts compared with the recurrent baseline.

Real-world archival footage. MgfrOFR produces visually cleaner restorations on real-world SRWOV clips without ground-truth supervision at test time.

Experimental Analysis

Ablation. MGFE and RATR provide complementary improvements over the baseline.	Sensitivity. Token/branch choices are stable around the default full-model setting.
Convergence. The full model preserves stable optimization behavior.	Branch diversity. RATR reduces redundant branch behavior and encourages multi-granularity specialization.

TODO

• Add paper link.
• Add pretrained model download URLs and SHA256 checksums.
• Add final benchmark tables after metric provenance is frozen.
• Add qualitative examples and method figures.

Repository Structure

MgfrOFR/
├── mgfrofr/              # MgfrOFR architecture, model wrapper, and RATR loss
├── basicofr/             # Video restoration runtime components
├── basicsr/              # BasicSR-style training/testing infrastructure
├── options/
│   ├── train/            # Training and ablation configs
│   └── test/             # Synthetic and real-world evaluation configs
├── scripts/              # Train/test entry points
├── docs/                 # Dataset, model-zoo, and result notes
├── tests/                # Import/registry/config smoke test
├── CITATION.cff
├── LICENSE
├── requirements.txt
└── setup.py

Installation

1. Clone

git clone https://github.com/wscffaa/MgfrOFR.git
cd MgfrOFR

2. Create Environment

conda create -n mgfrofr python=3.10 -y
conda activate mgfrofr

Install PyTorch according to your CUDA version from the official PyTorch instructions. For example:

# Example only. Pick the command matching your CUDA/driver.
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

Then install the remaining dependencies and this repository:

pip install -r requirements.txt
pip install -e .

mamba-ssm and CUDA extensions can be environment-sensitive. If installation fails, install the PyTorch/CUDA-compatible wheel first and then rerun pip install -r requirements.txt.

Quick Check

Run the lightweight smoke test after installation:

python tests/test_mgfrofr_smoke.py

Expected output:

MgfrOFR import, registry, and config smoke test passed

This checks package imports, registry entries, and YAML config parsing. It does not require pretrained restoration weights.

Dataset Preparation

This repository does not redistribute REDS, SRWOV, or third-party datasets. Please download datasets from their official sources and arrange them as follows:

datasets/OldFilmRestoration/
├── train/
│   └── REDS/train_sharp/
├── val/
│   ├── miniREDS/
│   │   ├── GT/val_sharp/
│   │   └── LQ/val_sharp_degraded/
│   ├── REDS/
│   │   ├── GT/val_sharp/
│   │   └── LQ/val_sharp_degraded/
│   └── RealFilm/MambaOFR_SRWOV/
└── noise_data/

See docs/DATASETS.md for more notes.

Model Zoo

Weights are not committed to git. Put downloaded models under pretrained/:

pretrained/
├── mambaofr_reds_16_net_g_best.pth
├── mgfrofr_a3_full.pth
└── mgfrofr_a0_mambaofr.pth

Model	Config	Download	Notes
MgfrOFR-A3 full	options/test/mgfrofr_a3_synthetic.yml	To be released	Full MGFE + RATR model
MgfrOFR-A0 baseline	options/train/mgfrofr_a0_mambaofr.yml	To be released	Baseline reference
MambaOFR warm start	options/train/mgfrofr_a3_full.yml	To be released	Used for warm-start training

See docs/MODEL_ZOO.md. Add checksums before publishing pretrained weights.

Evaluation

Synthetic REDS-style Evaluation

python scripts/test.py -opt options/test/mgfrofr_a3_synthetic.yml

Outputs are saved under results/.

Real-world Old-film Evaluation

python scripts/test.py -opt options/test/mgfrofr_a3_realworld.yml

The real-world config saves restored frames and does not require ground-truth frames.

Training

Full MgfrOFR training:

python scripts/train.py -opt options/train/mgfrofr_a3_full.yml

Ablation configs:

Config	Purpose
options/train/mgfrofr_a0_mambaofr.yml	recurrent baseline
options/train/mgfrofr_a1_mgfe.yml	baseline + MGFE
options/train/mgfrofr_a2_ratr.yml	baseline + RATR
options/train/mgfrofr_a3_full.yml	MGFE + RATR full model

Training logs and checkpoints are written to experiments/ and tb_logger/, both ignored by git.

Results

Final paper metrics will be added after the evaluation protocol, checkpoint hashes, and metric provenance are frozen. See docs/RESULTS.md for the intended reporting fields.

When reporting results, please include:

• checkpoint file and hash;
• config file;
• dataset split;
• metric implementation;
• exact command;
• date and environment.

Citation

If this repository helps your research, please cite the paper. The final BibTeX will be updated after publication.

@article{cai2026mgfrofr,
  title={Multi-Granularity Feature Decomposition with Ranking-Aware Regularization for Old Film Restoration},
  author={Cai, Feifan},
  year={2026}
}

A machine-readable citation file is provided in CITATION.cff.

Acknowledgements

This repository follows the BasicSR-style restoration pipeline and includes BasicOFR/BasicSR runtime components required by MgfrOFR. MgfrOFR also builds on the old-film restoration setting and recurrent restoration baseline popularized by prior works such as MambaOFR and earlier old-film restoration frameworks. Please cite the relevant upstream projects and datasets when using this code.

License

This project is released under the Apache License 2.0.