cs.CVApr 3, 2026
SaveThe Eleventh NTIRE 2026 Efficient Super-Resolution Challenge Report
Abstract
This paper reviews the NTIRE 2026 challenge on efficient single-image super-resolution with a focus on the proposed solutions and results. The aim of this challenge is to devise a network that reduces one or several aspects, such as runtime, parameters, and FLOPs, while maintaining PSNR of around 26.90 dB on the DIV2K_LSDIR_valid dataset, and 26.99 dB on the DIV2K_LSDIR_test dataset. The challenge had 95 registered participants, and 15 teams made valid submissions. They gauge the state-of-the-art results for efficient single-image super-resolution.
Figures & tables
| Teams | PSNR [dB] | Runtime [ms] | #Params | FLOPs | Sub-Track Scores | Main-Track | ||||||
| Val. | Test | Val. | Test | Ave. | [M] | [G] | Runtime | #Params | FLOPs | Overall Score | Ranking | |
| XiaomiMM | (1) | (5) | (6) | 1 | ||||||||
| BOE_AIoT | (3) | (6) | (7) | 2 | ||||||||
| PKDSR | (4) | (7) | (8) | 3 | ||||||||
| DISP | (2) | (9) | (9) | |||||||||
| VARH-AI | (5) | (4) | (4) | |||||||||
Table 1: Results of Ninth NTIRE 2026 Efficient SR Challenge. The performance of the solutions is compared thoroughly from three perspectives including the runtime, FLOPs, and the number of parameters. The underscript numbers associated with each metric score denote the ranking of the solution in terms of that metric. For runtime, “Val.” is the runtime averaged on DIV2K_LSDIR_valid validation set. “Test” is the runtime averaged on a test set with 200 images from DIV2K_LSDIR_test set, respectively. “Ave.” is averaged on the validation and test datasets. “#Params” is the total number of parameters of a model. “FLOPs” denotes the floating point operations. Main Track combines all three evaluation metrics. The ranking for the main track is based on the score calculated via Eq. 2 , and the ranking for other sub-tracks is based on the score of each metric via Eq. 1 . Please note that this is not a challenge for PSNR improvement. The “validation/testing PSNR” is not ranked. For all the scores, the lower, the better.
Figure 1: Team XiaomiMM : SPANV2 overall architecture. The near-pixel branch (top) provides a pixel-repeat upsampling prior, while five SPABV2 blocks (bottom) extract deep features. The two paths are concatenated (80 ch) and fused by depthwise-separable convolution before PixelShuffle .
Figure 2: Team XiaomiMM : span_attn_op fuses the attention convolution, element-wise addition, and element-wise multiplication into a single CUDA kernel, eliminating redundant DRAM round-trips.
Figure 3: Team ZenoSR : The overall architecture of Adaptive Calibration Network(ACN)
Figure 4: Team CUIT_HTT : Overall architecture of the proposed MambaGate-SR.
Figure 5: Team HAESR : The architecture of HAESR
Figure 6: Team IN2GM : Overview of the proposed RFDN_SPAN architecture.
Figure 7: Team Sunflower : The network architecture of HFENet
Figure 8: Team XuptSR : The whole framework of Variance-Guided Spatial-Channel Context Interaction Network (VSCINet)
Figure 9: Team XuptSR : The details of each component. SCAB: Spatial Channel Attention Block; GSA: Gated Self-Attention; GDFN: Gated Depthwise Feed-Forward Network; GCAB:Global Channel Attention Block; VCSA: Variance-guided Contextual Spatial Attention; MSConv: Multi-scale Convolution; MSRB: Multi-scale Residual Block
Figure 10: Team WMESR : Overview of DWMamba.
Figure 11: Team VARH-AI : The proposed model architecture overview.
| Stage | Patch Size | Batch | Iterations | FFT Weight |
| 1 | 64 | 200K | 0 | |
| 2 | 64 | 200K | ||
| 3 | 64 | 200K | ||
| 4 | 64 | 200K | 0 |
Table 2: DSCLoRA multi-stage curriculum schedule.
Figure 12: Team XSR : Network architecture of the proposed AMCANet.
Figure 13: Team Just Try : Overall structure of ERRN2
Figure 14: Team MDAP : SAFMN-Deep15 overall architecture. The deep feature extraction path consists of 15 sequential SAFM blocks (utilizing multi-scale spatial pooling and GELU gating). The output is refined and added via a global residual before being finally upsampled by PixelShuffle( ).
More in this topic
Image Super-Resolution
Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.
See all 248 papers