ECHO_block4

ECHO_block4 is the DCD (Direct Conditional Distillation) stage distilled student model with a block length of 4. It achieves coherent report generation in a single forward pass per block, offering up to 8× inference speedup over multi-step baselines while maintaining high clinical quality.

ECHO (Efficient Chest X-ray Report Generation with One-step Block Diffusion) is a discrete diffusion vision–language model for automated chest X-ray report generation. DCD constructs non-factorized supervision from on-policy teacher trajectories, enabling coherent single-step decoding that was previously unachievable in discrete diffusion models.

Model Details

Usage

git clone https://github.com/midea-ai/ECHO.git
cd ECHO
pip install transformers==4.55.4

# Single-step inference with ECHO_block4 (distilled)
python inference/generate_echo.py \
  --model_dir Midea-AIRC/ECHO_block4 \
  --image_path /path/to/chest_xray.jpg \
  --prompt_text "Review this chest X-ray and write a report. Use this format: Findings: {}, Impression: {}." \
  --block_length 4 \
  --denoising_steps 1

For Chinese prompts:

python inference/generate_echo.py \
  --model_dir Midea-AIRC/ECHO_block4 \
  --image_path /path/to/chest_xray.jpg \
  --prompt_text "这是一组胸部X光图像，请生成一份医学报告，包括所见和结论。以以下格式返回报告：所见：{} 结论：{}。" \
  --block_length 4 \
  --denoising_steps 1

Model Zoo

License

This model is released under the Midea Model License Agreement - Non-Commercial Use Version. Use for research, study, and personal non-commercial purposes only. Commercial use is strictly prohibited.

Citation

@misc{chen2026echoefficientchestxray,
      title={ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion}, 
      author={Lifeng Chen and Tianqi You and Hao Liu and Zhimin Bao and Jile Jiao and Xiao Han and Zhicai Ou and Tao Sun and Xiaofeng Mou and Xiaojie Jin and Yi Xu},
      year={2026},
      eprint={2604.09450},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.09450}, 
}

Contact

• Lifeng Chen, Beijing Jiaotong University (lfchen@bjtu.edu.cn)
• Hao Liu (Corresponding Author), AI Research Center, Midea Group (liuhao249@midea.com)