Research map / Label-efficient and robust learning / Recognition and long-tail
SLVR: Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation
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 SLVR, a super-light single-image super-resolution network built from a new residual linking mode (FDEL) that reduces low-level feature retention and a new blueprint controllable convolution (B2Conv) with an access control unit gating intra-kernel correlation.
Problem
Lightweight super-resolution models suffer from network inertia caused by feature addition in residual blocks, and BSConv applies blueprint scale transformation uniformly across all channels, introducing useless or disruptive features.
Approach
Replace the input-plus-output residual addition with FDEL (output minus input, i.e., the block's gain), and replace BSConv's uniform scale transformation with B2Conv, where a sigmoid-gated access control unit adaptively selects which depth-axis channels get blueprint scaling.
Key results
- SLVR achieves 28.85 dB PSNR on Set14 x4 (SLVR* trained on DF2K) with 485K parameters and 26.2G Mult-Adds.
- SLVR at x4 on Set14 reaches 28.78 dB (DIV2K-trained) versus NGswin 28.78 dB with roughly half the parameters (485K vs 1,019K).
- Replacing BSConv with B2Conv in RFDN improves Set5 x4 from 32.12 to 32.22 dB with about 63K extra parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SLVR* instead of the strongest listed baseline | lightweight SR · Set5 ×4 | PSNR: 32.33 -> 32.44 | Helped |
| Use SLVR* instead of the strongest listed baseline | lightweight SR · Set14 ×4 | PSNR: 28.78 -> 28.85 | Helped |
| Use SLVR* instead of the strongest listed baseline | lightweight SR · BSD100 ×4 | PSNR: 27.66 -> 27.70 | Helped |
| Use SLVR* instead of the strongest listed baseline | lightweight SR · Urban100 ×4 | PSNR: 26.45 -> 26.58 | Helped |
| Use SLVR* instead of the strongest listed baseline | lightweight SR · Manga109 ×4 | PSNR: 30.84 -> 31.10 | Helped |
| Add FDEL linking in place of addition links (ablation) | EDSR-baseline · Set5 ×4 | PSNR: 32.08 -> 32.12 | Helped |
| Add FDEL linking in place of addition links (ablation) | EDSR-baseline · Set14 ×4 | PSNR: 28.58 -> 28.61 | Helped |
| Add FDEL linking in place of addition links (ablation) | EDSR-baseline · Urban100 ×4 | PSNR: 25.98 -> 26.05 | Helped |
| Add B2Conv instead of BSConv (ablation) | RFDN · Set5 ×4 | PSNR: 32.12 -> 32.22 | Helped |
| Add B2Conv instead of BSConv (ablation) | RFDN · Set14 ×4 | PSNR: 28.59 -> 28.65 | Helped |
| Add B2Conv instead of BSConv (ablation) | RFDN · Urban100 ×4 | PSNR: 26.01 -> 26.06 | Helped |
| Add FDEL and B2Conv to the SLVR framework (ablation) | SLVR · Set5 ×4 | PSNR: 32.37 -> 32.40 | Helped |
Surprising findings
- Replacing the output of one ESDB block with an earlier one had almost no effect on reconstruction quality (PSNR drop under 0.005 dB), taken as evidence of network inertia.
- In the RFDN study, B2Conv with 261K params at x3 gives results close to ordinary convolution with 633K params on Set14 (30.36 vs 30.35 dB).
Limitations
- Gains from FDEL and B2Conv are small on several benchmarks (e.g., 0.01 dB differences in ablations).
- Comparisons to other SOTA lightweight models are mostly on bicubic-downsampled benchmarks; the ACU target values are only idealized as 0 or 1 and softened to (0,1).
Experimental setup
Datasets: DIV2K Flickr2K Set5 Set14 B100 Urban100 Manga109
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109
Metrics: PSNR SSIM parameters Mult-Adds
Baselines: IMDN LAPAR-A LatticeNet RLFN SRPN-Lite FMEN MLRN NGswin HPUN-M SAFMN OSFFNet BSRN
Builds on: BSConv BSRN ESDB / IMDN-style information distillation EDSR RFDN
Closest papers by meaning
- SuperPrimitive: Scene Reconstruction at a Primitive Level (CVPR 2024 · similarity 0.83)
- Low-Light Face Super-resolution via Illumination, Structure, and Texture Associated Representation (AAAI 2024 · similarity 0.83)
- OSFFNet: Omni-Stage Feature Fusion Network for Lightweight Image Super-Resolution (AAAI 2024 · similarity 0.83)
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.82)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.82)
- S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting (CVPR 2025 · similarity 0.82)
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.