Research map / Video / Video segmentation and tracking
Arbitrary-Scale Video Super-resolution Guided by Dynamic Context
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 DCGU, a temporal-context-guided filter-based upsampling module that predicts non-local sampling locations and content-dependent filter weights for efficient arbitrary-scale video super-resolution.
Problem
Arbitrary-scale video super-resolution: pixelshuffle upsampling cannot handle arbitrary scales, while implicit neural function methods are slow and filter-based methods have limited representation capacity.
Approach
Replaces fixed local sampling in filter-based upsampling with context-guided non-local sampling (multi-granularity search over LR grid via LSH-based global patch search and local pixel search with dynamic gating), and modulates filter weights with a normalized similarity between HR temporal context and LR features (temporal bilateral filter modulation).
Key results
- At x4 scale, BasicVSR++ DCGU reaches 32.57 dB PSNR on REDS4 versus 32.39 dB for BasicVSR++, at 75 ms vs 77 ms runtime.
- In-distribution arbitrary-scale results with BasicVSR++ backbone: DCGU reaches 32.57 dB at x4 vs 32.31 dB (LTE) at 75 ms vs 301 ms for LTE.
- Out-of-distribution x8 and x12 on Vimeo-90K-T: DCGU gets 27.46 dB and 25.46 dB, outperforming LTE (27.20, 25.25).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace pixelshuffle upsampling with DCGU | BasicVSR++ · REDS4, Vimeo-90K-T, and Vid4, x4 BI | PSNR: 32.39 -> 32.57; 37.79 -> 37.96; 27.79 -> 27.85 | Helped |
| replace LTE upsampling with DCGU | BasicVSR++ · REDS4, x2.0, x2.5, x3.0, x3.5, and x4.0 BI | PSNR: 38.10 -> 38.55; 6.67 -> 37.05; 34.92 -> 35.19; 33.46 -> 33.70; 32.31 -> 32.57 | Helped |
| replace LTE upsampling with DCGU | BasicVSR++ · REDS4, x8.0 and x12.0 BI | PSNR: 27.20 -> 27.46; 25.25 -> 25.46 | Helped |
| use bicubic-upsampled features as the HR context (ablation) | DCGU · REDS4, x4.0 BI | PSNR: 32.57 -> 32.21 | Hurt |
| use MetaSR to generate the HR context (ablation) | DCGU · REDS4, x4.0 BI | PSNR: 32.57 -> 32.52 | Hurt |
| add local pixel search (ablation) | DCGU · REDS4, x4.0 BI | PSNR: 32.27 -> 32.38 | Helped |
| add dynamic gating to local pixel search (ablation) | DCGU · REDS4, x4.0 BI | PSNR: 32.38 -> 32.45 | Helped |
| add global patch search (ablation) | DCGU with local pixel search and dynamic gating · REDS4, x4.0 BI | PSNR: 32.45 -> 32.57 | Helped |
| replace MGLS with DCN-predicted sampling locations (ablation) | DCGU · REDS4, x4.0 BI | PSNR: 32.57 -> 32.35 | Hurt |
| add negative L2 distance as range-filter modulation (ablation) | DCGU spatial filter · REDS4, x4.0 BI | PSNR: 32.48 -> 32.50 | Helped |
| add normalized similarity as range-filter modulation (ablation) | DCGU spatial filter · REDS4, x4.0 BI | PSNR: 32.48 -> 32.57 | Helped |
Surprising findings
- Plugging DCGU into BasicVSR adds parameters but still runs faster, because the filter prediction runs once per video and the pixelshuffle convolutions on HR features are removed.
Limitations
- Efficiency gains are measured relative to a BasicVSR++ backbone; upsampling runtime is only reported as backbone-plus-module total.
- Out-of-distribution evaluation is limited to a few scales (x8, x12) on one dataset in the main text.
Experimental setup
Datasets: REDS Vimeo-90K
Benchmarks: REDS4 Vimeo-90K-T Vid4 UDM10
Metrics: PSNR SSIM parameters (M) runtime (ms)
Baselines: MetaSR LIIF LTE BasicVSR BasicVSR++ ArbSR VideoINR
Builds on: MetaSR BasicVSR++ DCN locality-sensitive hashing non-local attention
Closest papers by meaning
- SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network (AAAI 2024 · similarity 0.90)
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution (CVPR 2025 · similarity 0.89)
- Deep Video Super-Resolution Network Using Dynamic Upsampling Filters Without Explicit Motion Compensation (CVPR 2018 · similarity 0.89)
- Enhancing Video Super-Resolution via Implicit Resampling-based Alignment (CVPR 2024 · similarity 0.87)
- DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution (NeurIPS 2025 · similarity 0.87)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 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.