NTIRE 2026 3D Content Super-Resolution Challenge Solutions

This repository contains the 2D front-end of our solution to the NTIRE 2026 3D Content Super-Resolution Challenge, covering:

• Track 1: Bicubic Degradation
• Track 2: Realistic Degradation

This repository focuses on the 2D front-end pipeline, including rendered image restoration and enhancement for downstream 3D content super-resolution.
For the 3D reconstruction part, please refer to our 3D repository:
3D-Super-Resolution

For the full pipeline, this front-end repository is also responsible for the required depth preprocessing step before 3DGS training.

Challenge Website

• Codabench Competition Track 1
• Codabench Competition Track 2

Model Checkpoints

We provide the pretrained checkpoints for SRFormerV2, MambaIRV2, and HAT-L here.

Track 1: Bicubic Degradation

Instruction

We use SRFormerV2 to perform super-resolution of rendered LR in Track1

Execution

# activate SRFormerV2
conda activate srformer

# SRFormerV2 inference
python models/SRFormer/basicsr/test.py -opt models/SRFormer/options/test/SRFormerV2/002_SRFormer_3DSR_from_pretrain_test_final.yml

Track2: Realistic Degradation

Instruction

We utilize three different models to jointly exploit branch-level stabilization and cross-model complementarity, and ensemble their outputs to produce more reliable restored views for the subsequent 3D reconstruction stage. HAT was pretrained on DIV2K and Flickr2K, we also finetuned HAT models on OST and LSDIR dataset.

Execution

# activate mambairv2
conda activate mambair

# MambaIRv2 inference
python models/MambaIR/basicsr/test.py -opt models/MambaIR/options/test/mambairv2/004_My_3DSR_MambaIRv2_RealSR_x4_tta_test.yml

# Rename the inference results
python models/MambaIR/rename.py \
"/path/to/track2/EastResearchAreas" \
"/path/to/track2/NorthAreas" \
--pattern "_mamba.png" \
--dest ".JPG"

# SRFormer
conda activate srformer

# SRFormer inference
python models/SRFormer/basicsr/test.py -opt models/SRFormer/options/test/SRFormerV2/002_SRFormer_3DSR_from_pretrain_real_test_final.yml

# Rename the inference results
python models/MambaIR/rename.py \
"/path/to/track2/EastResearchAreas" \
"/path/to/track2/NorthAreas" \
--pattern "_002_SRFormer_3DSR_from_pretrain_real_test_final.png" \
--dest ".JPG"

# HAT-L
conda activate HAT

# HAT-L inference
torchrun --nproc_per_node=1 \
  --nnodes=1 \
  --node_rank=0 \
  --master_addr=127.0.0.1 \
  --master_port=29511 \
  models/HAT/hat/test.py -opt models/HAT/options/test/NTIRE2026/004_HAT-L_3DSRx4_from_pretrain_real_test_final.yml \
  --launcher pytorch

# Rename the inference results
python models/MambaIR/rename.py \
"/path/to/track2/EastResearchAreas" \
"/path/to/track2/NorthAreas" \
--pattern "_hat.png" \
--dest ".JPG"
  
# Model Ensemble
# 0.01 MambaIR + 0.01 SRFormer + 0.98 HAT
conda activate HAT

# 1， Conduct model ensemble for EastResearchAreas
python models/HAT/ensemble.py \
--folder1 /path/to/MambaIR/track2/EastResearchAreas \
--folder2 /path/to/SRFormer/track2/EastResearchAreas \
--folder3 /path/to/HAT/track2/EastResearchAreas \
--weights 0.01 0.01 0.98 \
--output /path/to/track2/EastResearchAreas/rgb

# Conduct model ensemble for NorthAreas
python models/HAT/ensemble.py \
--folder1 /path/to/MambaIR/track2/NorthAreas \
--folder2 /path/to/SRFormer/track2/NorthAreas \
--folder3 /path/to/HAT/track2/NorthAreas \
--weights 0.01 0.01 0.98 \
--output /path/to/track2/NorthAreas/rgb

Depth

Depth preprocessing is shared by both tracks and should be completed before starting 3DGS training. We use cv2.INTER_LINEAR to upscale depth/*.png by x4.

This command writes the upscaled depth files to /path/to/xxx_depth_x4 by default while preserving the scene/depth directory structure.

You can also provide an explicit output directory:

python upscale_depth.py /path/to/xxx /path/to/output_depth_x4