Papers › Activating More Pixels in Image Super-Resolution Transformer

Activating More Pixels in Image Super-Resolution Transformer

9 May 2022CVPR 2023 1arXiv:2205.04437archive 2025-07-28

Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, Chao Dong

Transformer-based methods have shown impressive performance in low-level vision tasks, such as image super-resolution. However, we find that these networks can only utilize a limited spatial range of input information through attribution analysis. This implies that the potential of Transformer is still not fully exploited in existing networks. In order to activate more input pixels for better reconstruction, we propose a novel Hybrid Attention Transformer (HAT). It combines both channel attention and window-based self-attention schemes, thus making use of their complementary advantages of being able to utilize global statistics and strong local fitting capability. Moreover, to better aggregate the cross-window information, we introduce an overlapping cross-attention module to enhance the interaction between neighboring window features. In the training stage, we additionally adopt a same-task pre-training strategy to exploit the potential of the model for further improvement. Extensive experiments show the effectiveness of the proposed modules, and we further scale up the model to demonstrate that the performance of this task can be greatly improved. Our overall method significantly outperforms the state-of-the-art methods by more than 1dB. Codes and models are available at https://github.com/XPixelGroup/HAT.

PaperPDFConference PDFCode

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

chxy95/hat officialmentioned in papermentioned on GitHubpytorchApache-2.0 report
xpixelgroup/hat officialmentioned in papermentioned on GitHubpytorchApache-2.0 report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Image Super-ResolutionSuper-Resolution

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Image Super-Resolution BSD100 - 2x upscaling HAT-L PSNR 32.74 #7 of 30 Archive leaderboard report
Image Super-Resolution BSD100 - 2x upscaling HAT-L SSIM 0.9066 #7 of 30 Archive leaderboard report
Image Super-Resolution BSD100 - 2x upscaling HAT PSNR 32.69 #9 of 30 Archive leaderboard report
Image Super-Resolution BSD100 - 2x upscaling HAT SSIM 0.9060 #9 of 30 Archive leaderboard report
Image Super-Resolution BSD100 - 3x upscaling HAT-L PSNR 29.63 #3 of 21 Archive leaderboard report
Image Super-Resolution BSD100 - 3x upscaling HAT-L SSIM 0.8191 #3 of 21 Archive leaderboard report
Image Super-Resolution BSD100 - 3x upscaling HAT PSNR 29.59 #5 of 21 Archive leaderboard report
Image Super-Resolution BSD100 - 3x upscaling HAT SSIM 0.8177 #5 of 21 Archive leaderboard report
Image Super-Resolution BSD100 - 4x upscaling HAT-L PSNR 28.09 #4 of 71 Archive leaderboard report
Image Super-Resolution BSD100 - 4x upscaling HAT-L SSIM 0.7551 #4 of 71 Archive leaderboard report
Image Super-Resolution BSD100 - 4x upscaling HAT PSNR 28.05 #8 of 71 Archive leaderboard report
Image Super-Resolution BSD100 - 4x upscaling HAT SSIM 0.7534 #8 of 71 Archive leaderboard report
Image Super-Resolution Manga109 - 2x upscaling HAT-L PSNR 41.01 #4 of 21 Archive leaderboard report
Image Super-Resolution Manga109 - 2x upscaling HAT-L SSIM 0.9831 #4 of 21 Archive leaderboard report
Image Super-Resolution Manga109 - 2x upscaling HAT PSNR 40.71 #6 of 21 Archive leaderboard report
Image Super-Resolution Manga109 - 2x upscaling HAT SSIM 0.9819 #6 of 21 Archive leaderboard report
Image Super-Resolution Manga109 - 3x upscaling HAT-L PSNR 36.02 #3 of 17 Archive leaderboard report
Image Super-Resolution Manga109 - 3x upscaling HAT-L SSIM 0.9576 #3 of 17 Archive leaderboard report
Image Super-Resolution Manga109 - 3x upscaling HAT PSNR 35.84 #5 of 17 Archive leaderboard report
Image Super-Resolution Manga109 - 3x upscaling HAT SSIM 0.9567 #5 of 17 Archive leaderboard report
Image Super-Resolution Manga109 - 4x upscaling HAT-L PSNR 33.09 #4 of 50 Archive leaderboard report
Image Super-Resolution Manga109 - 4x upscaling HAT-L SSIM 0.9335 #4 of 50 Archive leaderboard report
Image Super-Resolution Manga109 - 4x upscaling HAT PSNR 32.87 #6 of 50 Archive leaderboard report
Image Super-Resolution Manga109 - 4x upscaling HAT SSIM 0.9319 #6 of 50 Archive leaderboard report
Image Super-Resolution Set14 - 2x upscaling HAT-L PSNR 35.29 #3 of 35 Archive leaderboard report
Image Super-Resolution Set14 - 2x upscaling HAT-L SSIM 0.9293 #3 of 35 Archive leaderboard report
Image Super-Resolution Set14 - 2x upscaling HAT PSNR 35.13 #6 of 35 Archive leaderboard report
Image Super-Resolution Set14 - 2x upscaling HAT SSIM 0.9282 #6 of 35 Archive leaderboard report
Image Super-Resolution Set14 - 3x upscaling HAT-L PSNR 31.47 #3 of 24 Archive leaderboard report
Image Super-Resolution Set14 - 3x upscaling HAT-L SSIM 0.8584 #3 of 24 Archive leaderboard report
Image Super-Resolution Set14 - 3x upscaling HAT PSNR 31.33 #5 of 24 Archive leaderboard report
Image Super-Resolution Set14 - 3x upscaling HAT SSIM 0.8576 #5 of 24 Archive leaderboard report
Image Super-Resolution Set14 - 4x upscaling HAT-L PSNR 29.47 #4 of 104 Archive leaderboard report
Image Super-Resolution Set14 - 4x upscaling HAT-L SSIM 0.8015 #4 of 104 Archive leaderboard report
Image Super-Resolution Set14 - 4x upscaling HAT PSNR 29.38 #7 of 104 Archive leaderboard report
Image Super-Resolution Set14 - 4x upscaling HAT SSIM 0.8001 #7 of 104 Archive leaderboard report
Image Super-Resolution Set5 - 2x upscaling HAT-L PSNR 38.91 #3 of 41 Archive leaderboard report
Image Super-Resolution Set5 - 2x upscaling HAT-L SSIM 0.9646 #3 of 41 Archive leaderboard report
Image Super-Resolution Set5 - 2x upscaling HAT PSNR 38.73 #6 of 41 Archive leaderboard report
Image Super-Resolution Set5 - 2x upscaling HAT SSIM 0.9637 #6 of 41 Archive leaderboard report
Image Super-Resolution Set5 - 3x upscaling HAT-L PSNR 35.28 #3 of 32 Archive leaderboard report
Image Super-Resolution Set5 - 3x upscaling HAT-L SSIM 0.9345 #3 of 32 Archive leaderboard report
Image Super-Resolution Set5 - 3x upscaling HAT PSNR 35.16 #8 of 32 Archive leaderboard report
Image Super-Resolution Set5 - 3x upscaling HAT SSIM 0.9335 #8 of 32 Archive leaderboard report
Image Super-Resolution Set5 - 4x upscaling HAT-L PSNR 33.30 #2 of 12 Archive leaderboard report
Image Super-Resolution Set5 - 4x upscaling HAT-L SSIM 0.9083 #2 of 12 Archive leaderboard report
Image Super-Resolution Urban100 - 2x upscaling HAT-L PSNR 35.09 #4 of 29 Archive leaderboard report
Image Super-Resolution Urban100 - 2x upscaling HAT-L SSIM 0.9505 #4 of 29 Archive leaderboard report
Image Super-Resolution Urban100 - 2x upscaling HAT PSNR 34.81 #7 of 29 Archive leaderboard report
Image Super-Resolution Urban100 - 2x upscaling HAT SSIM 0.9489 #7 of 29 Archive leaderboard report
Image Super-Resolution Urban100 - 3x upscaling HAT-L PSNR 30.92 #3 of 22 Archive leaderboard report
Image Super-Resolution Urban100 - 3x upscaling HAT-L SSIM 0.8981 #3 of 22 Archive leaderboard report
Image Super-Resolution Urban100 - 3x upscaling HAT PSNR 30.70 #5 of 22 Archive leaderboard report
Image Super-Resolution Urban100 - 3x upscaling HAT SSIM 0.8949 #5 of 22 Archive leaderboard report
Image Super-Resolution Urban100 - 4x upscaling HAT-L PSNR 28.60 #4 of 65 Archive leaderboard report
Image Super-Resolution Urban100 - 4x upscaling HAT-L SSIM 0.8498 #4 of 65 Archive leaderboard report
Image Super-Resolution Urban100 - 4x upscaling HAT PSNR 28.37 #7 of 65 Archive leaderboard report
Image Super-Resolution Urban100 - 4x upscaling HAT SSIM 0.8447 #7 of 65 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionBPEConcatenated Skip ConnectionDense ConnectionsDropoutLabel SmoothingLayer NormalizationLinear LayerMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections