Research map / Architectures and efficiency / Pruning and efficient tuning
Lightweight Image Super-Resolution via Flexible Meta Pruning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes flexible meta pruning (FMP), which jointly prunes channels (structured) and kernel weights (unstructured) of a lightweight SR network via a hypernetwork driven by channel vectors and weight indicators, plus a new lightweight SR baseline LSRB.
Problem
Lightweight image super-resolution models still contain substantial redundancy; existing compression methods (NAS, KD, channel pruning) are costly, need pretrained models, or neglect unstructured weight pruning.
Approach
A hypernetwork takes per-layer channel vectors (structured pruning of output/input channels) and weight indicators (unstructured pruning of kernel weights) and generates backbone weights, so both are sparsified jointly during training from scratch, with proximal gradient for channel vectors and SGD for weight indicators.
Key results
- FMP applied to LSRB achieves best PSNR/SSIM on all five benchmarks at ×2, ×3, ×4; e.g., Urban100 ×4 gains over second-best ASSLN (about 0.0047 SSIM gain at ×4).
- At ×4 on Urban100, LSRB-6-48 gets 25.62 dB vs RLFN 25.54 dB with faster inference (12 ms vs 19 ms on RTX 3090).
- On EDSR-8-128, FMP beats DHP channel pruning at all prune ratios (e.g., Urban100 25.96 vs 25.92 at 60% prune).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply flexible channel-and-weight pruning | LSRB · Set5 ×4 | PSNR: 32.29 -> 32.34 | Helped |
| apply flexible channel-and-weight pruning | LSRB · Set14 ×4 | PSNR: 28.69 -> 28.71 | Helped |
| apply flexible channel-and-weight pruning | LSRB · B100 ×4 | PSNR: 27.66 -> 27.67 | Helped |
| apply flexible channel-and-weight pruning | LSRB · Urban100 ×4 | PSNR: 26.27 -> 26.35 | Helped |
| apply flexible channel-and-weight pruning | LSRB · Manga109 ×4 | PSNR: 30.84 -> 30.90 | Helped |
| add weight pruning to channel pruning at 60% prune ratio (ablation) | EDSR-8-128 · Set5 ×4 | PSNR: 31.99 -> 32.16 | Helped |
| add weight pruning to channel pruning at 60% prune ratio (ablation) | EDSR-8-128 · Set14 ×4 | PSNR: 28.52 -> 28.60 | Helped |
| add weight pruning to channel pruning at 60% prune ratio (ablation) | EDSR-8-128 · B100 ×4 | PSNR: 27.53 -> 27.55 | Helped |
| add weight pruning to channel pruning at 60% prune ratio (ablation) | EDSR-8-128 · Urban100 ×4 | PSNR: 25.92 -> 25.96 | Helped |
| replace L2 regularization with L1 regularization for weight indicators (ablation) | EDSR-8-128 + FMP · Urban100 ×4 | PSNR: 25.81 -> 25.90 | Helped |
| replace weight decay with L1 regularization for weight indicators (ablation) | EDSR-8-128 + FMP · Urban100 ×4 | PSNR: 25.90 -> 25.90 | No clear effect |
| replace RLFN residual block with a simplified residual block (ablation) | LSRB-6-48 · Urban100 ×4 | inference time (ms): 19 -> 12 | Helped |
Surprising findings
- L1 norm and weight decay regularization on weight indicators outperform L2 norm for weight pruning in this setting.
Limitations
- Only applied to EDSR and the newly designed LSRB; not tested on other SR architectures such as SwinIR.
- Pruning experiments trained from scratch; no comparison under pretrained-model settings in the main ablations.
- Authors note lack of convincing pruning benchmarks for SR.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 B100 Urban100 Manga109 DIV2K
Metrics: PSNR SSIM Params FLOPs inference time
Baselines: CARN IMDN LatticeNet ASSLN DIPNet DHP MoreMNAS-A FALSR-A CARN+KD RLFN
Builds on: DHP RLFN EDSR ASSLN MetaPruning
Closest papers by meaning
- Learning Efficient Image Super-Resolution Networks via Structure-Regularized Pruning (ICLR 2022 · similarity 0.93)
- Aligned Structured Sparsity Learning for Efficient Image Super-Resolution (NeurIPS 2021 · similarity 0.93)
- Image Processing GNN: Breaking Rigidity in Super-Resolution (CVPR 2024 · similarity 0.88)
- Achieving Lightweight Super-Resolution for Real-Time Computer Graphics (AAAI 2025 · similarity 0.87)
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · similarity 0.87)
- Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution (AAAI 2025 · similarity 0.86)
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.