NTIRE 2026 Challenge on Mobile Real-World Image Super-Resolution @ CVPR 2026

About the Challenge

The challenge is part of the 11th NTIRE Workshop at CVPR 2026, which targets the real-world image super-resolution on mobile devices. Participants should recover a high‑resolution image from a single low‑resolution input that is 4 × smaller and with unknown degradations.

The evaluation consists of comparing the restored high-resolution images with the ground truth high-resolution images. To comprehensively assess the results, we employ evaluation metrics as follows:

• Inference Speed: We will benchmark the inference speed on the MediaTek Dimensity 8400 platform, using the inference speed of OSEDiff on this platform as the baseline. The input image size is $128\times 128$, and the ouput size is $512\times 512$. We define $t_{osediff}$ and $t_{curmodel}$ as the average inference time on single image using OSEDiff and current model, and the definition of speedup ratio is:

$$ Speedup=\frac{t_{osediff}}{t_{curmodel}} $$

• Perceptual Metrics: LPIPS, DISTS, NIQE, ManIQA, MUSIQ, and CLIP-IQA. To measure the super-resolution performance, we calculate the average weighted value of the six perceptual metrics. The input image size is arbitrary. The Score is defined as follows:

$$ \text{Score} = \left(1 - \text{LPIPS}\right) + \left(1 - \text{DISTS}\right) + \text{CLIPIQA} + \text{MANIQA} + \frac{\text{MUSIQ}}{100} + \max\left(0, \frac{10 - \text{NIQE}}{10}\right) $$

The final score of each participant is defined as follows:

$$ FinalScore=2^{Score}\cdot {Speedup}^{0.2} $$

Challenge results

• 16 valid submissions are ranked.
• Evaluation set: all scores are measured on the DIV2K‑val (100 images) with unknown degradations.
• Overall order: ranking depends on the $FinalScore$.

Certificates

The top three teams in this competition have been awarded NTIRE 2026 award certificates.

All certificates can be downloaded from Google Drive.

How to test the model?

1. git clone https://github.com/jiatongli2024/NTIRE2026_Mobile_RealWorld_ImageSR.git

2. Download the model weights from:

   • Google Drive

   Put the downloaded weights in the ./model_zoo folder.

3. Select the model you would like to test:

   CUDA_VISIBLE_DEVICES=0 python test.py --valid_dir [path to val data dir] --test_dir [path to test data dir] --save_dir [path to your save dir] --model_id 0

   • You can use either --valid_dir, or --test_dir, or both of them. Be sure the change the directories --valid_dir/--test_dir and --save_dir.
   • We provide a baseline (team00): DAT (default). Switch models (default is DAT) through commenting the code in test.py.

4. Some methods cannot be integrated into our codebase. We provide their instructions in the corresponding folder. If you still fail to test the model, please contact the team leaders. Their contact information is as follows:

   Index	Team	Leader	Email
   1	VIPSL	JiaHao Deng	1695185764djh@gmail.com
   2	Antman	Zhenzhong Chen	zzchen@whu.edu.cn
   3	SamsungAICamera	Yoonjin Im	yoonjin.im@samsung.com
   4	TODSR	Zihao Wang	wwzzhh@njust.edu.cn
   5	YuFans	Wei Zhou	weichow@u.nus.edu
   6	IMAG2006	Xinzhe Zhu	xzzhu@njust.edu.cn
   7	Super03	Runze Tian	Trz220765@mail.ustc.edu.cn
   8	VEPG	Congyu Wang	congyuwang@njust.edu.cn
   9	SnowVision	Choulhyouc Lee	iron.lee@snowcorp.com
   10	BVISR	Yuxuan Jiang	dd22654@bristol.ac.uk
   11	EIC-ECNU	Shaohui Lin	shlin@cs.ecnu.edu.cn
   12	NTR	Jiachen Tu	jtu9@illinois.edu
   13	NoReject	Yuqi Li	yuqili010602@gmail.com
   14	ACM_HCC	Shyang-En Weng	shyangenweng.cs13@nycu.edu.tw
   15	MDAP	Watchara Ruangsang	watchara.knot@gmail.com
   16	SFVision	Yuwen Pan	panyuwen@sz.tsinghua.edu.cn

How to eval images using IQA metrics?

Environments

conda create -n NTIRE-SR python=3.8
conda activate NTIRE-SR
pip install -r requirements.txt

Folder Structure

test_dir
├── HR
│   ├── 0901.png
│   ├── 0902.png
│   ├── ...
├── LQ
│   ├── 0901x4.png
│   ├── 0902x4.png
│   ├── ...
    
output_dir
├── 0901x4.png
├── 0902x4.png
├──...

Command to calculate metrics

python eval.py \
--output_folder "/path/to/your/output_dir" \
--target_folder "/path/to/test_dir/HR" \
--metrics_save_path "./IQA_results" \
--gpu_ids 0 \

The eval.py file accepts the following 4 parameters:

• output_folder: Path where the restored images are saved.
• target_folder: Path to the HR images in the test dataset. This is used to calculate FR-IQA metrics.
• metrics_save_path: Directory where the evaluation metrics will be saved.
• device: Computation devices. For multi-GPU setups, use the format 0,1,2,3.

License and Acknowledgement

This code repository is release under MIT License.