NTIRE 2025 Challenge on Image Super-Resolution (x4) @ CVPR 2025 -[BBox Team 15]

The souce files of the factsheet are available at factsheet.

How to test the team15 model?

Step1: load docker image

docker load --input bbox.tar BaiduYun:r6ji

Step2: download the pretrained model

• The pretrained models are available at model_zoo/team15_SMT/team15_SMT_HAT.txt and model_zoo/team15_SMT/team15_SMT_Mamba.txt.
• Download the two pretrained models and put them into model_zoo.

Step3: inference

1. For single GPU:

CUDA_VISIBLE_DEVICES=<gpu_id> torchrun --nproc_per_node=1 test.py --model_id 15 test_dir [path to test data dir] --save_dir [path to your save dir]

2. For multi-GPU:

CUDA_VISIBLE_DEVICES=<gpu_ids> torchrun --nproc_per_node=<num_gpus> test.py --model_id 15 test_dir [path to test data dir] --save_dir [path to your save dir]

If you are using a different number or type of GPUs, please adjust the nproc_per_node parameter accordingly to match your hardware configuration.

Note:🚨 Inference speed on a single GPU will be relatively slow. Specifically, we use 10× A100 GPUs for inference, which takes approximately 40 minutes. The extended inference time is due to the multi-window self-ensemble strategy applied to two models, along with the final model ensemble integration.