LR2Flow: Enhancing Low-resolution Image Representation Through Normalizing Flows

Official PyTorch implementation of LR2Flow, presented in:

Enhancing Low-resolution Image Representation Through Normalizing Flows Chenglong Bao, Tongyao Pang, Zuowei Shen, Dihan Zheng, and Yihang Zou arXiv preprint arXiv:2601.06834, 2026.

[Paper]

Pretrained models are available from Google Drive.

Please download the corresponding .pth files and place them in:

./experiments/pretrained_models/

Before training or testing, please configure the corresponding .yml files located in:

./codes/options/train/
./codes/options/test/

The dataset paths, pretrained model paths, and other experimental settings can be specified in these configuration files.

Run the following commands under the ./codes/ directory.

×2 image rescaling

python3 test_rescaling.py --opt options/test/test_rescaling_x2.yml

×4 image rescaling

python3 test_rescaling.py --opt options/test/test_rescaling_x4.yml

python3 test_compression.py --opt options/test/test_compression_x2.yml

Alternatively, to evaluate the compression performance in terms of bits per pixel (bpp), run:

python3 test_compression_bpp.py --opt options/test/test_compression_x2.yml

python3 test_denoising.py --opt options/test/test_denoising.yml

Run the following commands under the ./codes/ directory.

×2 image rescaling

python3 train_rescaling.py --opt options/train/train_rescaling_x2.yml

×4 image rescaling

python3 train_rescaling.py --opt options/train/train_rescaling_x4.yml

python3 train_compression.py --opt options/train/train_compression_x2.yml

python3 train_denoising.py --opt options/train/train_denoising.yml

This codebase is built upon HCFlow and is also inspired by IRN and BasicSR.