Restormer: Efficient Image Restoration Transformer

• Restormer is a Transformer architecture designed for high-resolution image restoration, efficiently modeling long-range dependencies with channel-wise attention.
• It features a U-Net-like hierarchical encoder-decoder with MDTA and GDFN modules that reduce complexity and enhance performance across various degradation tasks.
• The model sets state-of-the-art results in deraining, deblurring, and denoising, establishing a strong baseline and guiding subsequent research advancements.

Restormer, short for Restoration Transformer, is an efficient Transformer architecture for high-resolution image restoration that was introduced to retain the long-range modeling advantages of Transformers while avoiding the quadratic spatial cost of vanilla self-attention (Zamir et al., 2021). It is formulated as a U-Net-like hierarchical encoder-decoder Transformer for image-to-image restoration, and its defining architectural move is to replace token-wise spatial attention with channel-wise transposed attention, combined with a convolution-enhanced gated feed-forward module. In the original study, the model was evaluated on deraining, single-image motion deblurring, defocus deblurring, Gaussian denoising, and real image denoising, and reported state-of-the-art or near-state-of-the-art results across 16 benchmarks (Zamir et al., 2021).

1. Problem setting and design objective

Restormer was proposed against a specific background in low-level vision. CNN-based restorers were already strong at learning image priors, but their convolutional receptive fields remained limited and their filters were static at inference. Standard Transformers, by contrast, could model long-range dependencies and adapt to content, but their self-attention cost grew quadratically with spatial resolution, making them impractical for many restoration problems involving large images (Zamir et al., 2021).

The central design objective was therefore not merely to import a vision Transformer into restoration, but to construct a model that preserves global interaction while remaining feasible on large restoration inputs. Earlier restoration Transformers often reduced cost by restricting attention to windows or independent patches, which weakens global dependency modeling. Restormer instead keeps attention global in effect, but restructures it so that the complexity is linear with respect to spatial resolution (Zamir et al., 2021).

This design objective also explains why Restormer rapidly became a reference architecture in subsequent work. Later papers repeatedly use it either as a baseline, a teacher model, or a reconstruction backbone, which suggests that the architecture established a durable operating point between restoration quality, scalability, and architectural simplicity (Liu et al., 23 Jul 2025).

2. Core architecture

Restormer uses a 4-level symmetric encoder-decoder. Given a degraded image

$\mathbf{I}\in \mathbb{R}^{H\times W\times 3},$

a convolution first produces shallow features, which are then processed by a hierarchical encoder-decoder with skip connections and a high-resolution refinement stage. The output is predicted residually as

$\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$

In the default configuration, the number of blocks across levels is $[4,6,6,8]$, the number of channels is $[48,96,192,384]$, the MDTA heads are $[1,2,4,8]$, and the refinement stage has 4 blocks (Zamir et al., 2021).

The first defining block is multi-Dconv head transposed attention (MDTA). For a layer-normalized input $Y$, Restormer forms query, key, and value features by combining $1\times1$ point-wise convolutions with $3\times3$ depth-wise convolutions:

$$Q = W_d^Q W_p^Q Y,\quad K = W_d^K W_p^K Y,\quad V = W_d^V W_p^V Y.$$

Attention is then computed in channel space rather than over all spatial positions:

$$\hat{X} = W_p\,\textrm{Attention}(\hat{Q}, \hat{K}, \hat{V}) + X,$$

with

$\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$0

Here $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$1 is a learnable scaling parameter. This transposed construction avoids a $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$2 attention matrix and is the primary reason the model scales to large images (Zamir et al., 2021).

The second defining block is the gated-Dconv feed-forward network (GDFN). Instead of a standard Transformer feed-forward layer, Restormer uses a gated, depth-wise-convolutional formulation:

$\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$3

where

$\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$4

Here $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$5 is GELU and $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$6 is element-wise multiplication. This introduces both local spatial inductive bias and content-adaptive gating into the feed-forward path (Zamir et al., 2021).

Architecturally, Restormer is therefore neither a plain ViT nor a CNN with attention add-ons. It is a restoration-specific Transformer in which both the attention block and the feed-forward block are re-engineered around the demands of dense image recovery.

3. Training protocol and empirical profile

The original training setup uses AdamW with $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$7, $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$8, weight decay $\hat{\mathbf{I}} = \mathbf{I} + \mathbf{R}.$9, $[4,6,6,8]$0 loss, 300K iterations, an initial learning rate of $[4,6,6,8]$1, cosine annealing down to $[4,6,6,8]$2, and horizontal and vertical flips as augmentation (Zamir et al., 2021). For denoising, a bias-free version of Restormer is used. A notable training strategy is progressive learning, in which patch size increases as

$[4,6,6,8]$3

while batch size is reduced accordingly; the reported effect is a measurable PSNR gain over fixed-size patch training (Zamir et al., 2021).

Across restoration tasks, the original paper reports a broad empirical profile. On image deraining, Restormer improves the average PSNR over SPAIR by 1.05 dB, with gains as high as 2.06 dB on Rain100L. On single-image motion deblurring, averaged across GoPro, HIDE, RealBlur-R, and RealBlur-J, it improves over MPRNet by 0.26 dB and over MIMO-UNet+ by 0.47 dB; relative to MPRNet, it uses 81% fewer FLOPs, and relative to IPT, it has 4.4× fewer parameters and is 29× faster while delivering better quality (Zamir et al., 2021).

On defocus deblurring, Restormer is reported to improve the combined DPDD result by about 0.6 dB over IFAN and by 1.01 dB over Uformer. On Gaussian denoising, it achieves the best PSNR on all listed grayscale datasets and noise levels, and on Urban100 at $[4,6,6,8]$4 it improves by 0.37 dB over DRUNet and 0.31 dB over SwinIR in grayscale denoising, while in color denoising the gain is about 0.41 dB over DRUNet and 0.2 dB over SwinIR. On real image denoising, it is described as the only method reported to surpass 40 dB PSNR on both SIDD and DND (Zamir et al., 2021).

A plausible implication is that Restormer’s importance lies not in one isolated benchmark result, but in the fact that the same architectural template remained competitive across tasks with different degradation statistics, ranging from rain streak removal to real sensor noise.

4. Architectural descendants and comparative developments

A substantial later literature treats Restormer as a strong baseline but also interrogates its specific inductive biases. DiNAT-IR is explicitly framed as a response to a weakness in Restormer-style restoration Transformers: while Restormer’s channel-wise self-attention is efficient and effective, it may overlook fine local artifacts and pixel-level structure. DiNAT-IR therefore keeps the same four-stage U-Net backbone as Restormer, replaces the attention primitive with Dilated Neighborhood Attention (DiNA), and adds a channel-aware module (CAM) to reintroduce global context. On GoPro and HIDE, it reports 33.80 dB / 0.967 SSIM and 31.57 dB / 0.945 SSIM, compared with Restormer’s 32.92 dB / 0.961 and 31.22 dB / 0.942; on SIDD and Rain100L, however, it is slightly below Restormer in PSNR while remaining close in SSIM (Liu et al., 23 Jul 2025).

Other work has pursued more heterogeneous attention designs. DART describes Restormer as strong but limited by channel-complexity growth and a single-attention bias, and proposes a multi-attention design combining LongIR attention, local and global attention, plus feature- and positional-dimension attention. Its reported DART-B model has 4.5M parameters, versus 26.13M for Restormer, while slightly exceeding Restormer on SIDD (40.10 dB / 0.961 versus 40.02 dB / 0.960) and on DND (40.06 dB / 0.957 versus 40.03 dB / 0.956) (Wen et al., 2024).

A separate line of work argues that Restormer’s architecture is not always the primary bottleneck. ConStyle v2 keeps the Restormer backbone unchanged and adds a frozen prompt generator that turns it into an all-in-one image restoration model. In the reported 27-benchmark average, Original Restormer is listed at 6.40 / 0.0506, while ConStyle v2 Restormer reaches 27.60 / 0.8352 (Fan et al., 2024). This suggests that prompt conditioning and training protocol can sometimes dominate architecture redesign.

There are also direct efficiency-oriented modifications. One study on high-resolution motion deblurring reports an 18.4% reduction in parameter count, about 30% fewer layers, lower memory footprint, richer augmentation, and a new frequency-domain loss while preserving competitive quality on RealBlur-R, RealBlur-J, and UHDM (Akmaral et al., 30 Jan 2025). A later edge-efficiency study uses Restormer as the transformer teacher and baseline, then distills its blocks into Mamba-style state-space surrogates; on a Snapdragon 8 Elite CPU, the Restormer baseline requires 10119.52 ms for inference, whereas ENS-discovered hybrids reduce this to 2973 ms for deblurring, 5816 ms for deraining, and 8666 ms for denoising while maintaining competitive restoration quality (Miriyala et al., 4 May 2026).

Taken together, these developments do not displace Restormer so much as clarify its position: it is a stable architectural reference point against which locality-aware, data-centric, prompt-guided, and edge-efficient alternatives are defined.

5. Domain-specific integrations

Restormer has also been embedded into domain-specific systems in which the degradation model is not a generic blur or noise process but part of a larger sensing pipeline.

Domain	Restormer’s role	Reported result
Hyperbolic metalens imaging	Learned post-processing module for severe off-axis aberrations	Wide FOV imaging over 54°; about 50 ms per image (Yeo et al., 29 Jul 2025)
Atmospheric turbulence mitigation	Core reconstruction backbone with a NIMA residual branch	98.53% average accuracy on text-pattern reconstruction (Liu et al., 2022)
QR code motion deblurring	Baseline transformed into EG-Restormer and routed within ADNet	90.00% DR at 0.737 s for ADNet (Li et al., 14 Oct 2025)
Universal MRI reconstruction	Learned proximal/reconstruction operator in each SDUM cascade	Backbone ablation: 32.090 PSNR / 0.881 SSIM for Restormer (Wang et al., 19 Dec 2025)

In the hyperbolic metalens study, Restormer is used as a direct image-to-image restoration model for a camera whose limitation is not on-axis focal quality but strong off-axis aberrations. The network is trained in a reference-free manner using simulated blurred images generated with the eigenPSF method, and the reported system enables “high-quality imaging over a wide FOV of 54°” on experimentally captured scenes under diverse lighting conditions (Yeo et al., 29 Jul 2025).

In the UG$[4,6,6,8]$5+ atmospheric turbulence mitigation report, Restormer is the main reconstruction backbone and is combined with a NIMA-based image quality assessment module through residual fusion. The model processes 20 randomly selected frames from a 100-frame turbulence sequence, uses $[4,6,6,8]$6 sliding windows with stride 100, and the final system achieves 1st place on the final leaderboard with 98.53% average accuracy on the reconstruction result of the text patterns (Liu et al., 2022).

For QR code motion deblurring, the emphasis shifts from perceptual quality to successful decoding. The paper introduces EG-Restormer, which augments Restormer with an Edge-Guided Attention Block, and then combines it with a lightweight path in ADNet. With GoPro + QRData fine-tuning, Restormer reports 88.67% DR, while EG-Restormer reaches 90.00% DR; ADNet preserves 90.00% DR while reducing average inference time from 0.910 s to 0.737 s (Li et al., 14 Oct 2025).

In MRI, Restormer functions not as a standalone restorer but as the image-domain reconstructor inside an unrolled optimization framework. In SDUM, every cascade alternates coil-sensitivity estimation, sampling-aware weighted data consistency, and a Restormer-based reconstruction step. Under a controlled backbone ablation at $[4,6,6,8]$7, the Restormer option reports 32.090 PSNR / 0.881 SSIM, ahead of U-Net, DiT, and TAU under the same setup (Wang et al., 19 Dec 2025).

6. Limitations, robustness, and continuing debates

A common misconception is that Restormer’s strong benchmark accuracy makes it uniformly reliable at inference. Follow-up work shows that this is not the case. One line of analysis revisits global information aggregation and argues that restoration models trained on cropped patches but tested on full-resolution images suffer a train-test inconsistency. Applied to Restormer, the proposed Test-time Local Converter (TLC) yields Restormer-Local, improving GoPro from 32.92 dB to 33.57 dB and HIDE from 31.22 dB to 31.49 dB without retraining or fine-tuning (Chu et al., 2021).

A stronger critique concerns adversarial robustness. On GoPro deblurring, clean-image performance for Restormer is reported as 31.99 PSNR / 0.9635 SSIM, but under CosPGD, 20 iters it drops to 7.59 / 0.1548. The same study reports that FGSM adversarial training raises Restormer’s attacked performance to 21.58 / 0.7317, and argues that Restormer has better robust generalization potential than simplified descendants such as Baseline and NAFNet, even though it remains vulnerable without explicit hardening (Agnihotri et al., 2023).

Stability in the Lipschitz sense is another limitation. A later denoising study uses Restormer as a strong unconstrained baseline and argues that, under mild perturbations such as JPEG compression, Restormer’s output can deviate much more than that of a provably contractive denoiser. In standalone denoising, Restormer is usually best or near-best among unconstrained models, but in PnP restoration the proposed contractive denoiser often surpasses it substantially, which indicates that raw restoration accuracy and convergence-certified inverse-problem behavior are distinct criteria (Shukla et al., 27 Mar 2026).

There is also a practical misconception that architectural novelty is always the dominant source of improvement. A data-centric denoising study starts from the public Restormer $[4,6,6,8]$8 pretrained model, keeps the backbone unchanged, expands the training corpus to 143,679 images from seven public sources, and applies $[4,6,6,8]$9 geometric self-ensemble. Under a unified 100-image validation protocol, the final method improves over the public Restormer baseline by up to +3.3662 dB PSNR, while the retained TLC-style local inference wrapper is reported to have negligible quantitative effect in that setting (Chang et al., 13 Apr 2026).

These debates locate Restormer’s historical significance more precisely. It is an influential restoration Transformer and a strong empirical baseline, but later work shows that its efficiency is context-dependent, its global aggregation can be inference-sensitive, its clean-image accuracy does not imply robustness, and substantial gains can also come from data, prompting, or physics-based integration rather than from replacing MDTA or GDFN outright.