This repository contains the official implementation for ABCD introduced in the following paper:

ABCD : Arbitrary Bit-wise Coefficient for De-quantization (CVPR 2023)

Our ABCD reconstruct randomly quantized images with single training!

Our code is based on Ubuntu 20.04, pytorch 1.10.0, CUDA 11.3 (NVIDIA RTX 3090 24GB, sm86) and python 3.6.

For enviorment setup, we recommend to use conda for installation:

conda env create --file environment.yaml
conda activate abcd

Our train and valid sets follows the prior work Punnappurath et. al., sampled from MIT-Adove FiveK and Sintel dataset.

For the benchmark part in Tab 1, we tested Kodak, TESTIMAGES1200 (\B01C00 folder), and ESPLv2 datasets.

The basic train code is :

python train.py --config configs/train_ABCD/SomeConfigs.yaml --gpu 0

We provide configurations of our main paper

EDSR-ABCD : configs/train_ABCD/train_EDSR-ABCD.yaml

RDN-ABCD : configs/train_ABCD/train_RDN-ABCD.yaml

SwinIR-ABCD : configs/train_ABCD/train_SwinIR-ABCD.yaml

EDSR-baseline_ABCD : configs/train_ABCD/train_EDSR-baseline_ABCD.yaml

If you want to modify some configuration (e.g. the range of input bit-depth) modify .yaml files and run

python train.py --config configs/train_ABCD/FancyConfiguration.yaml --gpu 0

We provide pre-trained models of our ABCD. Test our models with test codes below.

Check the training time and required the number of GPU for training.

We recommend trying EDSR-baseline_ABCD since it is lighter than models above.

To test our model, run the code below :

python test.py --config configs/test_ABCD/abcd_test-16bits.yaml --model save/PATHS/MODEL.pth --LBD 4 --HBD 8 --gpu 0

16-bit image dataset : /configs/test_ABCD/abcd_test-16bits.yaml (e.g. MIT-5K, TESTIMAGES1200)

8-bit image dataset : /configs/test_ABCD/abcd_test-8bits.yaml (e.g. Kodak)

If you want to test another labels, you may change flag 'LBD' and 'HBD' to test your model.

(e.g. 3-bits to 12-bits --LBD 3 --HBD 12, respectively)

For SwinIR based ABCD, test code needs additional flag --window 8

You can save results with using the flag --save 1.

This code is built on LIIF,LTE and SwinIR. We thank the authors for sharing their codes.