NTIRE 2026 Image Denoising (σ=50) — Team Noice (SRX) — 6th Place

Official solution of Team Noice for the NTIRE 2026 Image Denoising Challenge (σ=50) @ CVPRW 2026, finishing 6th place on the final leaderboard.

We propose SRX, an ensemble pipeline that fuses three independently fine-tuned restoration backbones — Restormer, XFormer, and SCUNet — combined with tiled inference and 8-fold geometric test-time augmentation (TTA), followed by a weighted linear ensemble.

Method Overview

SRX is built upon an ensemble of three state-of-the-art transformer- and CNN-based restoration networks. Each model is independently fine-tuned on a combined high-quality dataset and fused via a weighted ensemble strategy at inference time. Tiled inference and 8-fold D4 TTA are applied to each branch to boost performance and handle arbitrarily large inputs efficiently.

Figure 1: Overview of the proposed SRX pipeline. Each of the three branches independently denoises the noisy input via tiled inference + 8-fold TTA; the outputs are then fused with a weighted linear combination to produce the final denoised image.

Branch 1 — Restormer (σ=50 fine-tuned)

Hierarchical encoder–decoder transformer using transposed (channel-wise) multi-head self-attention instead of spatial attention, making it tractable for high-resolution images.

• dim = 48
• num_blocks = [4, 6, 6, 8]
• num_refinement_blocks = 4
• heads = [1, 2, 4, 8]
• FFN expansion factor: 2.66
• BiasFree LayerNorm (required for the Gaussian denoising variant)
• Initialized from the official gaussian_color_denoising_sigma50.pth checkpoint

Branch 2 — XFormer (σ=50 fine-tuned)

Hybrid X-shaped architecture combining spatial window-based self-attention and channel-wise attention, giving receptive fields complementary to Restormer.

• dim = 48
• num_blocks = [2, 4, 4]
• spatial_num_blocks = [2, 4, 4, 6]
• num_refinement_blocks = 4
• heads = [1, 2, 4, 8]
• Window size: 16 × 16 at each scale
• Initialized from the publicly released denoising weights, then fine-tuned

Branch 3 — SCUNet (σ=50 fine-tuned)

Swin-Conv UNet integrating Swin Transformer blocks and convolutional residual blocks in a symmetric encoder–decoder structure.

• config = [4, 4, 4, 4, 4, 4, 4] (seven-stage depth)
• dim = 64
• Initialized from the official scunet_color_50.pth checkpoint, then fine-tuned

Tiled Inference

Since inputs can be arbitrarily large, all three branches use overlapping tile-based inference:

• Tile size: 128 × 128
• Overlap: 32 px per border (stride = 128 - 32 = 96)
• Overlapping regions are accumulated and averaged with a pixel-wise weight map to eliminate seam artefacts

Test-Time Augmentation (8-fold, D4 group)

Each branch averages predictions over all 8 symmetries of the dihedral group D4:

1. Identity
2. Horizontal flip
3. Vertical flip
4. Horizontal + vertical flip (180° rotation)
5. Transpose
6. Transpose + horizontal flip
7. Transpose + vertical flip
8. Transpose + horizontal + vertical flip

Each augmented view is passed through the model (with tiled inference), the inverse transform is applied to the output, and all 8 results are averaged pixel-wise.

Weighted Ensemble

The final image is a weighted linear combination of the TTA-averaged outputs of the three branches:

Î = 0.3 · Î_SCUNet + 0.5 · Î_XFormer + 0.2 · Î_Restormer

Weights were tuned empirically on the CodaBench validation leaderboard — XFormer receives the highest weight as it showed the strongest standalone performance on the validation split.

Training Details

Data. Combined clean-image corpus, with σ=50 Gaussian noise synthesized on-the-fly (no external noisy images used):

• LSDIR — 2,000 sampled images
• DIV2K — 800 training images

Augmentation. Random 64×64 crop (bilinear upscale if the source is smaller) → on-the-fly AWGN (σ=50) → random horizontal/vertical flips → random 90° rotations.

Loss. ℓ1 (mean absolute error) between the restored patch and the clean ground truth.

Optimizer. AdamW, lr = 1e-4, β₁ = 0.9, β₂ = 0.999.

Schedule. Batch size 4, patch size 64×64, 40 epochs per model, mixed precision (AMP + GradScaler).

Partial freezing (Restormer only). The three encoder stages (encoder_level1/2/3) are frozen during fine-tuning; only the bottleneck, decoder, and refinement blocks are updated. This avoids overfitting the already well-trained low-level features given the relatively small fine-tuning set.

Each of the three backbones is fine-tuned independently under this same protocol.

Setup

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

Download Weights

Download the fine-tuned checkpoints and place all three files in model_zoo/.

Note: double check the individual links below before sharing this repo — all three currently point to the same file ID. Grab the correct file for each checkpoint name from the folder link, or replace these with the correct per-file share links.

Folder (all checkpoints): https://drive.google.com/drive/folders/1jxCEg71H4Clrm-ERoDg9O6gP9fjbxNPB?usp=sharing

File	Branch	Link
team08_RestormerFT.pth	Restormer	Download
team08_XformerFT.pth	XFormer	Download
team08_ScunetFT.pth	SCUNet	Download

Run Inference

python test_demo.py --data_dir ./data --save_dir ./results --model_id 8

This runs the full SRX pipeline: tiled inference + 8-fold TTA for each of the three branches, followed by the weighted ensemble, on every image in --data_dir, saving results to --save_dir.

Results

• Development phase entry ID: 599941
• Validation phase entry ID: 625178
• Final ranking: 6th place (NTIRE 2026 Image Denoising, σ=50 track)
• Quantitative PSNR/SSIM results are available on the official NTIRE 2026 CodaBench leaderboard.

Citation

If you find this work useful, please consider citing the official challenge report:

@article{ntire2026dn50,
  title   = {The Third Challenge on Image Denoising at NTIRE 2026: Methods and Results},
  journal = {arXiv preprint arXiv:2606.16031},
  year    = {2026}
}

Acknowledgements

Our SRX pipeline builds on the official implementations of Restormer, XFormer, and SCUNet. We thank the NTIRE 2026 organizers for hosting a well-structured and fair evaluation environment on CodaBench.