QDM: Quadtree-Based Region-Adaptive Sparse Diffusion Models for Efficient Image Super-Resolution

If you've found QDM useful for your research or projects, please show your support by ⭐ in this repo. Thanks!

News: This paper was accepted to CVPR 2026 Findings on March 18, 2026.

---

Deep learning-based super-resolution (SR) methods often perform pixel-wise computations uniformly across entire images, even in homogeneous regions where high-resolution refinement is redundant. We propose the Quadtree Diffusion Model (QDM), a region-adaptive diffusion framework that leverages a quadtree structure to selectively enhance detail-rich regions while reducing computations in homogeneous areas. By guiding the diffusion with a quadtree derived from the low-quality input, QDM identifies key regions—represented by leaf nodes—where fine detail is essential and applies minimal refinement elsewhere. This mask-guided, two-stream architecture adaptively balances quality and efficiency, producing high-fidelity outputs with low computational redundancy. Experiments demonstrate QDM’s effectiveness in high-resolution SR tasks across diverse image types, particularly in medical imaging (e.g., CT scans), where large homogeneous regions are prevalent. Furthermore, QDM outperforms or is comparable to state-of-the-art SR methods on standard benchmarks while significantly reducing computational costs, highlighting its efficiency and suitability for resource-limited environments.

---

Update

• 2026.03.18: Paper accepted to CVPR 2026 Findings.
• 2025.11.18: Released a new arXiv version with tumor region reconstruction and real-world SR results. Refer to the paper for details. Use print_roi_metrics.py to replicate the tumor reconstruction results. Access results for all methods here. Updated real-world SR with Gaussian-weighted patch-level aggregation as per this reference in utils/util_image.py.
• 2025.03.18: Release codes & pretrained checkpoints, and update README.
• 2025.03.14: Create this repo.

Requirements

• More detail (See requirements.txt) A suitable conda environment named quadtree_diffusion can be created and activated with:

conda create -n quadtree_diffusion python=3.10
conda activate quadtree_diffusion
pip install -r requirements.txt

Examples

Real-World Image Super-Resolution

Medical Image Super-Resolution

Fast Testing Guide

Download Pretrained Checkpoints

First-Stage Models (Autoencoders)

1. ​Real-world SR Task: Download Link
2. ​Medical SR Task: Download Link

​Note: Place the downloaded models in the weights directory.

QDM-L Checkpoints

We provide pretrained checkpoints for the QDM-L model for the following tasks:

• Real-world SR X4
• Medical SR X4
• Medical SR X8

​Note: Ensure all downloaded weights are placed in the weights directory.

Inference

🚀 Multi-GPU Acceleration

If you have multiple GPUs available, you can accelerate the inference process using the following command:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --standalone --nproc_per_node=8 --nnodes=1 inference.py \
  -i [Input Directory or Image] \
  -o [Output Dir] \
  --seed [Seed] \
  --chop_bs [Chopping Batch Size] \
  --chop_size [Chopping Size] \
  --cfg_path [Config Path] \
  --ckpt_path [Checkpoint Path] \
  --distributed

💻 Single-GPU Execution

python inference.py \
  -i [Input Directory or Image] \
  -o [Output Dir] \
  --seed [Seed] \
  --chop_bs [Chopping Batch Size] \
  --chop_size [Chopping Size] \
  --cfg_path [Config Path] \
  --ckpt_path [Checkpoint Path]

🔧Configuration Tips

• When processing very large images, you can adjust --chop_bs to balance efficiency and memory usage.
• We provide multiple configuration files for different tasks in the configs/inference directory. ​Make sure to select the appropriate configuration file for your specific task.
• You can add --process argument to output the mask guided diffusion process demonstrated in the paper.

Training

Preparing Stage

This repository supports two super-resolution (SR) tasks: Real-World SR and Medical CT SR. Follow the steps below to prepare the necessary training and testing datasets.

Real-World SR Task

We integrate training data from six established benchmarks:

• LSDIR – Access Dataset
• DIV2K – Access Dataset
• DIV8K – Access Dataset
• OutdoorSceneTraining – Access Dataset
• Flicker2K – Access Dataset
• FFHQ Subset – A curated selection of 10,000 facial images from the FFHQ dataset

Preprocessing Steps

• Filtering OutdoorSceneTraining:
  Filter out images with spatial dimensions smaller than 512 pixels. Update the directory path inside the script as needed, then run:

  python scripts/filter_images.py

• Synthetic LSDIR_TEST:
  Download the pre-synthesized LSDIR_TEST dataset from this link or generate your own by running:

  python scripts/prepare_lsdir_test.py

Medical CT SR Task

For the medical CT super-resolution task, we utilize clinical CT scans from two well-established segmentation challenges: HaN-Seg and SegRap2023. Download the datasets using the following links:

• Training Set: Download
• Medx4 Testing Set: Download
• Medx8 Testing Set: Download

Training Scripts

You can start your training process via running:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --standalone --nproc_per_node=8 --nnodes=1 main.py --cfg_path [Config Path] --save_dir [Logging Folder]

We provide multiple configuration files for different tasks in the configs/train directory. ​

Citations

Please consider citing our paper in your publications if it helps. Here is the bibtex:

@misc{yang2025qdmquadtreebasedregionadaptivesparse,
      title={QDM: Quadtree-Based Region-Adaptive Sparse Diffusion Models for Efficient Image Super-Resolution}, 
      author={Donglin Yang and Paul Vicol and Xiaojuan Qi and Renjie Liao and Xiaofan Zhang},
      year={2025},
      eprint={2503.12015},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.12015}, 
}

License

This project is licensed under MIT License. Redistribution and use should follow this license.

Acknowledgement

This project is primarily based on ResShift and LDM. We also adopt Real-ESRGAN to synthesize the LR/HR pairs. We design QDM mainly based on DiT. Thanks for their awesome works.

Contact

If you have any questions, please feel free to contact me via ydlin718@gmail.com.