Neural Video Compression with Context Modulation [CVPR 2025]
February 22, 2026 · View on GitHub
Neural Video Compression with Context Modulation [CVPR 2025]
Chuanbo Tang, Zhuoyuan Li, Yifan Bian, Li Li, Dong Liu
📌Overview
Our Deep Context Modulation for Video Compression (DCMVC) significantly advances the performance of Neural Video Codecs (NVCs). DCMVC is proposed to generate high-quality temporal context exploiting the reference information in both pixel and feature domain.
- Flow Orientation: It enables our DCMVC to generate additional oriented temporal context from the reference frame.
- Context Compensation: It eliminates the irrelevant propagated information to ensure better context modeling.
:bar_chart: Experimental Results
Main Results
Results comparison (BD-Rate and RD curve) for PSNR. The Intra Period is 32 with 96 frames. The anchor is VTM-13.2 LDB
Visualizations
- Our DCMVC enables better temporal context modeling.
Installation
This implementation of DCMVC is based on DCVC-DC and CompressAI. Please refer to them for more information.
1. Install the dependencies
conda create -n $YOUR_PY38_ENV_NAME python=3.8
conda activate $YOUR_PY38_ENV_NAME
conda install pytorch==1.10.0 torchvision==0.11.0 cudatoolkit=11.3 -c pytorch
pip install pytorch_ssim scipy matplotlib tqdm bd-metric pillow pybind11
2. Prepare test datasets
For testing the RGB sequences, we use FFmpeg to convert the original YUV 420 data to RGB data.
A recommended structure of the test dataset is like:
test_datasets/
├── HEVC_B/
│ ├── BQTerrace_1920x1080_60/
│ │ ├── im00001.png
│ │ ├── im00002.png
│ │ ├── im00003.png
│ │ └── ...
│ ├── BasketballDrive_1920x1080_50/
│ │ ├── im00001.png
│ │ ├── im00002.png
│ │ ├── im00003.png
│ │ └── ...
│ └── ...
├── HEVC_C/
│ └── ... (like HEVC_B)
└── HEVC_D/
└── ... (like HEVC_C)
3. Compile the arithmetic coder
If you need real bitstream writing, please compile the arithmetic coder using the following commands.
On Windows
cd src
mkdir build
cd build
conda activate $YOUR_PY38_ENV_NAME
cmake ../cpp -G "Visual Studio 16 2019" -A x64
cmake --build . --config Release
On Linux
sudo apt-get install cmake g++
cd src
mkdir build
cd build
conda activate $YOUR_PY38_ENV_NAME
cmake ../cpp -DCMAKE_BUILD_TYPE=Release
make -j
:rocket: Usage
1. Evaluation
Run the following command to evaluate the model and generate a JSON file that contains test results.
python test.py --rate_num 4 --test_config ./dataset_config_example_rgb.json --cuda 1 --worker 1 --output_path output.json --i_frame_model_path ./ckpt/cvpr2023_image_psnr.pth.tar --p_frame_model_path ./ckpt/dcmvc_p_frame.pth.tar
- We use the same Intra model as DCVC-DC.
cvpr2023_image_psnr.pth.tarcan be downloaded from DCVC-DC. - Our
dcmvc_p_frame.pth.tarcan be downloaded from CVPR2025-DCMVC.
Our model supports variable bitrate. Set different i_frame_q_indexes and p_frame_q_indexes to evaluate different bitrates.
:book: Citation
If this repo helped you, a ⭐ star or citation would make my day!
@InProceedings{tang2025neural,
author = {Tang, Chuanbo and Li, Zhuoyuan and Bian, Yifan and Li, Li and Liu, Dong},
title = {Neural Video Compression with Context Modulation},
booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
month = {June},
year = {2025},
pages = {12553--12563}
}
:email: Contact
If you have any questions, please contact me:
License
This work is licensed under MIT license.
Acknowledgement
Our work is implemented based on DCVC-DC and CompressAI.