Research map / Image generation / Image restoration
IRNeXt: Rethinking Convolutional Network Design for Image Restoration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-scale module (MSM) with 3x3 conv per branch to baseline (ablation) | U-shaped CNN baseline · GoPro ablation | PSNR: +0.23 dB (31.23 to 31.46) | Helped |
| replace convolution with local attention module (LAM) without filter modulation (ablation) | MSM-equipped baseline · GoPro ablation | PSNR: +0.07 dB (31.46 to 31.53) | Helped |
| add filter modulation (FM) to LAM (ablation) | LAM without FM · GoPro ablation | PSNR: +0.11 dB (31.53 to 31.64) | Helped |
| use three MSM branches (downsampling 8,4,2) instead of one branch (ablation) | MSM · GoPro ablation | PSNR: up to +0.41 dB over baseline with three branches | Helped |
| replace Softmax with Tanh for attention weights in LAM (ablation) | LAM · GoPro ablation | PSNR: +0.14 dB over Softmax (31.50 to 31.64) | Helped |
| use average pooling instead of max pooling or depthwise convolution for MSM downsampling (ablation) | MSM · GoPro ablation | PSNR: 31.64 vs 31.46 (max pooling) and 31.51 (convolution) | Helped |
| use IRNeXt in place of Restormer on DPDD | defocus deblurring · DPDD combined | PSNR: +0.32 dB (25.98 to 26.30) | Helped |
| use IRNeXt in place of DehazeFormer-L on dehazing | SOTS-Indoor dehazing · SOTS-Indoor | PSNR: +1.16 dB (40.05 to 41.21) | Helped |
| use IRNeXt in place of Stripformer and Restormer on GoPro deblurring | motion deblurring · GoPro | PSNR: +0.08 dB over Stripformer, +0.24 dB over Restormer, with ~4x faster inference | Helped |
Closest papers by meaning
- Restormer: Efficient Transformer for High-Resolution Image Restoration (CVPR 2022 · similarity 0.88)
- Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections (NeurIPS 2016 · similarity 0.88)
- Omni-Kernel Network for Image Restoration (AAAI 2024 · similarity 0.87)
- Efficient Visual State Space Model for Image Deblurring (CVPR 2025 · similarity 0.86)
- DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 2024 · similarity 0.85)
- MR-VNet: Media Restoration using Volterra Networks (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.