All-in-One Mural Restoration with Prompt-Guided Residual Diffusion

Installation

1. Download source code and dataset:

   • git clone https://github.com/CZY-Code/PGRDiff.git
   • Download the datasets
     • DUNHUANG
     • muralv2

2. Pip install dependencies:

   conda env create -f install.yaml
   

Dataset Preparation

Unzip and move dataset into ROOT

Directory structure of dataset

├── PGRDiff
│   ├── code
│   ├── DUNHUANG
│   │   ├── train
│   │   ├── test
│   ├── muralv2
│   │   ├── images
│   │   ├── masks
│   ├── install.yaml
│   ├── README.md

Training

cd ./code
python train.py

or

accelerate launch train.py

Evaluation

cd ./code
python metric.py

Pre-trained Models

├── code                     
├── DUNHUANG
│   ├── train         
│   ├── test
├──muralv2
│   ├── images
│   ├── masks
├── install.yaml
├── README.md

Training

cd ./code
python train.py

or

accelerate launch train.py

Evaluation

cd ./code
python metric.py

Pre-trained Models

• Download the weights of trained models
  • Weight for DUNHUANG
  • Weight for muralv2
• Move the weights into the folder ./code/results/sample/

Acknowledgement

This implementation is based on / inspired by:

• RDDM
• IR-SDE
• LRDiff
• StrDiffusion