CoD

<h2 align="center">CoD: A Diffusion Foundation Model for Image Compression</h2> <p align="center"> <a href="https://huggingface.co/papers/2511.18706"><img src="https://img.shields.io/badge/Paper-HF%20Paper%20Page-blue.svg" alt="Paper"></a> <a href="https://arxiv.org/abs/2511.18706"><img src="https://img.shields.io/badge/arXiv-2511.18706-b31b1b.svg" alt="arXiv"></a> <a href="https://github.com/microsoft/GenCodec/tree/main/CoD"><img src="https://img.shields.io/badge/Code-GitHub-blue.svg" alt="GitHub"></a> </p>

CoD (Compression-oriented Diffusion) is the first diffusion foundation model designed and trained from scratch specifically for image compression. It enables end-to-end optimization of both compression and generation.

Authors

Zhaoyang Jia, Zihan Zheng, Naifu Xue, Jiahao Li, Bin Li, Zongyu Guo, Xiaoyi Zhang, Houqiang Li, Yan Lu

Key Advantages

• High compression efficiency: Replaces Stable Diffusion in downstream codecs (like DiffC) to achieve SOTA results, especially at ultra-low bitrates (e.g., 0.0039 bpp).
• Low-cost and reproducible training: 300$\times$ faster training than Stable Diffusion ($\sim$ 20 vs. $\sim$ 6,250 A100 GPU days) on entirely open image-only datasets.
• Architecture: Features a lightweight condition encoder for image-native features, a VQ information bottleneck for compact bitstreams, and a Diffusion Transformer (DiT) for reconstruction.

Available Models

Base CoD Models (cod/)

One-Step CoD (finetuned_one_step_cod/)

Single forward pass, better performance, wider bitrates.

CoD as Perceptual Loss (perceptual_loss_illm_dec/)

Performance

Metrics evaluated on Kodak (512x512):

Note: CoD (latent) at 0.0039 bpp uses --cfg 1.25. CoD (latent, 64-bit) uses --cfg 3.0.

Quick Start

Installation

git clone https://github.com/microsoft/GenCodec.git
cd GenCodec/CoD
pip install -r requirements.txt

Download Checkpoints

# Download base CoD models
huggingface-cli download zhaoyangjia/CoD --include "cod/*" --local-dir ./pretrained/CoD

# Download one-step models
huggingface-cli download zhaoyangjia/CoD --include "finetuned_one_step_cod/*" --local-dir ./pretrained/CoD

# Download perceptual loss models
huggingface-cli download zhaoyangjia/CoD --include "perceptual_loss_illm_dec/*" --local-dir ./pretrained/CoD

# Download a specific model
huggingface-cli download zhaoyangjia/CoD cod/CoD_pixel_vpred.pt cod/CoD_pixel_vpred.yaml --local-dir ./pretrained/CoD

# Download everything
huggingface-cli download zhaoyangjia/CoD --local-dir ./pretrained/CoD

Base CoD Inference

python -m cod.inference evaluate \
    --ckpt ./pretrained/CoD/cod/CoD_pixel_vpred.pt \
    --config ./pretrained/CoD/cod/CoD_pixel_vpred.yaml \
    --input <image_dir> --output <recon_dir> \
    --step 25 --cfg 3.0 --sampler adam2

# For latent model, use --cfg 1.25
python -m cod.inference evaluate \
    --ckpt ./pretrained/CoD/cod/CoD_latent_vpred.pt \
    --config ./pretrained/CoD/cod/CoD_latent_vpred.yaml \
    --input <image_dir> --output <recon_dir> \
    --step 25 --cfg 1.25 --sampler adam2

One-Step CoD Inference

python -m downstream.finetuned_one_step_cod evaluate \
    --ckpt ./pretrained/CoD/finetuned_one_step_cod/bpp_0_0039.pt \
    --config ./pretrained/CoD/finetuned_one_step_cod/bpp_0_0039.yaml \
    --input <image_dir> --output <recon_dir>

Perceptual Loss Inference

Requires NeuralCompression (installed automatically via torch.hub).

python -m downstream.perceptual_loss_inference \
    --ckpt ./pretrained/CoD/perceptual_loss_illm_dec/msillm_quality_1.pt \
    --quality 1 \
    --input <image_dir> --output <recon_dir>

Citation

@inproceedings{jia2025cod,
    title     = {CoD: A Diffusion Foundation Model for Image Compression},
    author    = {Jia, Zhaoyang and Zheng, Zihan and Xue, Naifu and Li, Jiahao and Li, Bin and Guo, Zongyu and Zhang, Xiaoyi and Li, Houqiang and Lu, Yan},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2026}
}

License

MIT