[ECCV 2024] Pairwise Distance Distillation

Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution

Yuehan Zhang¹, Seungjun Lee², Angela Yao¹
National University of Singapore¹, Korea University²

📝PDD

We address the unsupervised RWSR for a targeted real-world degradation. We study from a distillation perspective and introduce a novel Pairwise Distance Distillation framework. Through our framework, a model specialized in synthetic degradation adapts to target real-world degradations by distilling intra- and inter-model distances across the specialized model and an auxiliary generalized model.

Our method, as a learning framework, can be applied to off-the-shelf generalist models, e.g., RealESRGAN, and improve their performance in a real-world domain!

🎯 TODOS

~~Complete code; release installation and training instructions~~
~~Public the repository~~
~~Debug environment settings~~
~~Release model weights~~
Update links to datasets
Update visual results

🖼️ Visual Examples

RealESRGAN + PDD | RealESRAGN

👓 Key Features

We tackle the unsupervised SR for a given real-world dataset through a distillation perspective:

we combine the knowledge from a Generalist (blind generalized model) and a Specialist (optimized for specific synthetic degradation);
we perform the distillation for distances between features of predictions, rather than features themselves;
only the Specialist is updated by gradients, and the full optimization consists of unsupervised distillation and supervised loss on synthetic data.

The distillation is based on the consistency of intra- and inter-model distances. We refer to the paper for explorations that establish these consistencies 😃.

We provide Static and EMA configurations for our method. The latter requires a single pretrained model and exhibits better performance.

🔨 Installation

Setup Environment

Python >= 3.9
PyTorch > 2.0
mmcv (recommend mmcv==2.1.0)
MMEngine
mmagic

Install PDD

# Clone the repository
git clone https://github.com/Yuehan717/PDD.git

# Navigate into the repository
cd PDD

# Install dependencies
pip install -r requirements.txt

If you only want to know the implementation, the main components of our method can be found in srgan_ref_model.py and losses.py 😃.

👉 Datasets & Model Weights

Datasets

Our method requires two sets of data for training:

Paired synthetic data: We use ground-truth images in DF2K and create LRs on the fly.
LR real-world data: In the paper, we experiment with RealSR, DRealSR, and NTRIE20.

Model Weights

[Here]

👉 Usage

# training instructions

### Please replace the path in .yml files with our own path.

## Distributed Training (recommend 2 gpus)
PYTHONPATH="./:${PYTHONPATH}" \
torchrun --nproc_per_node=[number_of_gpus] --master_port=[port] basicsr/train_mix.py \
-opt options/train/ESRGAN/[name_of_script].yml --launcher pytorch

## One-GPU Training
PYTHONPATH="./:${PYTHONPATH}" \
python basicsr/train_mix.py -opt options/train/ESRGAN/[name_of_script].yml

# Inference
python inference_SISR.py --input_path [path_to_input_folder] \
--save_path [path_to_result_folder] --model [esrgan/bsrgan] \
--model_path [path_to_model_weights] --color_corr

Some Tips

For EMA configuration, the hyperparameter domain_opt_warmup_iters (in [name_of_script].yml) can not be zero. It is to escape from the zero-distance state.
The color correction is implemented in reference script inference_SISR.py. The argument --color_corr calls this function.
The ratio between weights of intra- and inter-model distillation terms, as discussed in the paper, will influence the outputs' bias of fidelity and perception scores. For different real-world data, this ratio may need rounds of tuning to achieve the best balance.
The evaluation in the paper is based on IQA-PyTorch.

👏 Acknowledgement

The code is based on BasicSR. Thanks to their great contribution to the area!