🐱 DiT-IC Model Card

Source code is available at Github.

Model Description

• Developed by: NJU VisionLab
• Model type: Diffusion Transformer based image compression model
• Model size: 1.049 B parameters
• Resolution: Supports arbitrary resolution images

This model performs the diffusion process in a 32× latent space to reduce memory usage and accelerate inference.

The model is based on the pretrained text-to-image generative model SANA-600M.

Quick Start

For training and inference scripts, please visit our GitHub Repository.

Limitations

• The model may fail to reconstruct tiny text at low bitrates.
• Fingers and other fine structures may not be generated properly.

Citation

If you use this model, please cite:

@inproceedings{shi2026ditic,
  title={DiT-IC: Aligned Diffusion Transformer for Efficient Image Compression},
  author={Shi Junqi, Lu Ming, Li Xingchen, Ke Anle, Zhang Ruiqi and Ma Zhan},
  booktitle={CVPR},
  year={2026}
}