Research map / Architectures and efficiency / RNNs and normalization
WaveMixSR-V2: Enhancing Super-resolution with Higher Efficiency (Student Abstract)
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.
Enhances the WaveMixSR super-resolution model by replacing transposed convolution with PixelShuffle and adding a multi-stage 2x-block design for 4x SR, reporting SOTA on BSD100 with fewer parameters and higher throughput.
Problem
Single image super-resolution with high efficiency; prior WaveMixSR used a single non-parametric upsampling layer for scales above 2x and transposed convolutions that are parameter-heavy and costly.
Approach
Replace transposed convolution inside WaveMix blocks with PixelShuffle followed by a convolution, and turn higher-scale SR (4x) into a cascade of resolution-doubling 2x SR blocks instead of one interpolation-based upsampling step.
Key results
- WaveMixSR-V2 reaches state-of-the-art on BSD100 for super-resolution, per the abstract.
- WaveMixSR-V2 has 0.7M parameters and 25.6G multi-adds for 4x SR on 64x64 input, versus 1.7M and 25.8G for WaveMixSR, 11.8M/49.6G for SwinIR and 20.8M/103.7G for HAT.
- Inference throughput rises from 53.7 fps (WaveMixSR) to 82.6 fps (WaveMixSR-V2); inference latency drops from 18.6 ms to 12.1 ms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace transposed convolution with PixelShuffle followed by convolution | WaveMixSR · 4× SR, 64×64 input patch | #Params.: 1.7 M -> 0.7 M | Helped |
| Replace transposed convolution with PixelShuffle followed by convolution | WaveMixSR · 4× SR, 64×64 input patch | #Multi-Adds.: 25.8 G -> 25.6 G | Helped |
| Replace transposed convolution with PixelShuffle followed by convolution | WaveMixSR · Training | Latency: 22.8 -> 19.6 | Helped |
| Replace transposed convolution with PixelShuffle followed by convolution | WaveMixSR · Training | Throughput: 43.8 -> 50.8 | Helped |
| Replace transposed convolution with PixelShuffle followed by convolution | WaveMixSR · Inference | Latency: 18.6 -> 12.1 | Helped |
| Replace transposed convolution with PixelShuffle followed by convolution | WaveMixSR · Inference | Throughput: 53.7 -> 82.6 | Helped |
| Use a multi-stage design with successive 2× SR blocks instead of single-stage upsampling | WaveMixSR · 4× SR | super-resolution performance | Helped |
| Use the WaveMixSR-V2 architecture | WaveMixSR · BSD100 | PSNR | Helped |
| Use the WaveMixSR-V2 architecture | WaveMixSR · BSD100 | SSIM | Helped |
Limitations
- Multi-adds for WaveMixSR-V2 (25.6G) are nearly identical to WaveMixSR (25.8G), so computational savings are marginal.
- Evaluation appears limited to BSD100 in the provided text; other datasets are referenced only in general terms.
Experimental setup
Datasets: BSD100
Benchmarks: BSD100
Metrics: PSNR SSIM parameters multi-adds latency throughput
Baselines: SwinIR HAT WaveMixSR SwinFIR
Builds on: WaveMixSR WaveMix PixelShuffle
Closest papers by meaning
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.85)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.85)
- Continuous Optical Zooming: A Benchmark for Arbitrary-Scale Image Super-Resolution in Real World (CVPR 2024 · similarity 0.85)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.85)
- CAMixerSR: Only Details Need More "Attention" (CVPR 2024 · similarity 0.85)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · 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.