One-Step Diffusion-Based Image Compression with Semantic Distillation

NeurIPS 2025

Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, Yan Lu

👍 More Works

• DLF: Extreme Image Compression with Dual-generative Latent Fusion (ICCV 2025 Highlight)

📝 Abstract

While recent diffusion-based generative image codecs have shown impressive performance, their iterative sampling process introduces unpleasant latency. In this work, we revisit the design of a diffusion-based codec and argue that multi-step sampling is not necessary for generative compression. Based on this insight, we propose OneDC, a One-step Diffusion-based generative image Codec—that integrates a latent compression module with a one-step diffusion generator. Recognizing the critical role of semantic guidance in one-step diffusion, we propose using the hyperprior as a semantic signal, overcoming the limitations of text prompts in representing complex visual content. To further enhance the semantic capability of the hyperprior, we introduce a semantic distillation mechanism that transfers knowledge from a pretrained generative tokenizer to the hyperprior codec. Additionally, we adopt a hybrid pixel- and latent-domain optimization to jointly enhance both reconstruction fidelity and perceptual realism. Extensive experiments demonstrate that OneDC achieves SOTA perceptual quality even with one-step generation, offering over 39% bitrate reduction and 20×faster decoding compared to prior multi-step diffusion-based codecs. Project: https://onedc-codec.github.io/

⬆️ (Top) Multi-step sampling is not essential for image compression, One-Step is enough. (Bottom) Visual examples.

⬆️ OneDC can compress images to text-level size (a 768x768 image with 0.24KB), but the reconstruction still retains strong semantic consistency and original spatial details. (a) Text prompts (from GPT-4o) struggle to capture complex visual semantics, and existing text-to-image models have limited generation fidelity. (b) Hyperprior guidance yields more faithful reconstructions. (c) Semantic distillation further improves object-level accuracy.

⬆️ Framework overview.

⬆️ Quantitative evaluation.

💿 Installation

1. Create environment & install dependencies

conda create -n onedc python=3.10
conda activate onedc

Then, install PyTorch 2.5.0 manually according to your CUDA version. Finally, install required packages:

pip install -r requirements.txt

Notes:

• You may need to downgrade pip to v24.0.

2. Build the entropy coder

sudo apt-get install cmake g++
cd src
mkdir build
cd build
conda activate $YOUR_PY38_ENV_NAME
cmake ../cpp -DCMAKE_BUILD_TYPE=Release[Debug]
make -j

💻 Inference

Running coding process:

1. Download our model weights (lambda 12.2~0.6 model)
2. Run inference script:

cd src
python inference.py \
    --config_path config_inference.yaml \
    --checkpoint_path [checkpoint path] \
    --eval_image_path [your image folder] \
    --output_path [output folder] \

• For 0.0034 bpp model (exlow_bpp0034 model), use inference script in src/models/sd15_onedc_codec_z_only:

cd src
python models/sd15_onedc_codec_z_only/inference.py \
    --config_path config_inference.yaml \
    --checkpoint_path [checkpoint path] \
    --eval_image_path [your image folder] \
    --output_path [output folder] \

Evaluating quality:

cd src
python test_quality.py \
    --ref [your image folder] \
    --recon [recon image folder] \
    --fid_patch_size 256 \
    --fid_patch_num 2 \
    --output_path [result output folder] \
    --output_name [result output name]

🚀 Train

1. Prepare pretrained models: Pretrained one-step generator and VQGAN Tokenizer

2. Prepare dataset

• Download commoncatalog-cc-by-nd by script:

  cd src
  python get_cc_dataset.py -d [your download folder]

  Notes:

  • You may need to use "export HF_DATASETS_CACHE=" to change the HF cache dir, according to your disk space.

  The

  [your download folder]/common-canvas___commoncatalog-cc-by-nd/default/0.0.0/3991ff88ebf48e0435ec8d044d2f4b159f4f716e

  contains downloaded data file, e.g., "commoncatalog-cc-by-nd-train-00000-of-10447.arrow" and "dataset_info.json".

• Prepare your validation dataset.

  We use COCO 2017 dataset. You need to prepare annotations/captions_val2017.json and val2017 image set.

• Update the training config with the pretrained model and dataset path.

3. Stage I training

   cd src
   accelerate launch --config_file ddp_configs/ddp_4A100.yaml \
     models/sd15_onedc_codec_stage1/train_sd15_codec_stage1.py \
     --config_path models/sd15_onedc_codec_stage1/configs/config_sd15_onedc_lmbda4.6_stage1_lr5e-5.yaml \

4. Stage II training

   cd src
   accelerate launch --config_file ddp_configs/ddp_4A100.yaml \
     models/sd15_onedc_codec_stage2/train_sd15_codec_stage2.py \
     --config_path models/sd15_onedc_codec_stage2/configs/config_sd15_onedc_lmbda4.6_stage2_lr1e-6.yaml \

Notes:

• Adjust the lambda strategy in config to reach your target bitrate.
• For better results, you may need to manually decrease learning rate and continue training.

🥰 Acknowledgement

We sincerely thank the following outstanding works, which greatly inspired and supported our research:

• Improved Distribution Matching Distillation for Fast Image Synthesis
• MaskGIT: Masked Generative Image Transformer
• DCVC family

📕 Citation

If you find our work inspiring, please cite:

@inproceedings{xue2025one,
  title={One-Step Diffusion-Based Image Compression with Semantic Distillation},
  author={Naifu Xue and Zhaoyang Jia and Jiahao Li and Bin Li and Yuan Zhang and Yan Lu},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
}

⚖️ License

This project as a whole is licensed under CC BY-NC-SA 4.0.

It incorporates third-party components:

• DMD2: CC BY-NC-SA 4.0 (folder: src/modules/dmd)
• VQGAN: Apache License 2.0 (folder: src/modules/vqgan)
• DCVC: MIT License (file: src/modules/dcvc.py; folder: src/modules/entropy)