Research map / Image generation / Super-resolution
Continuous Optical Zooming: A Benchmark for Arbitrary-Scale Image Super-Resolution in Real World
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.
Introduces COZ, a real-world continuous optical zoom dataset captured with an automated motorized lens rig and SIFT-based alignment, plus the LMI MLP-mixer/meta-learning implicit network for arbitrary-scale super-resolution.
Problem
Arbitrary-scale image super-resolution models are trained on simulated bicubic-downsampled data, which fails to capture real-world degradation and yields unnatural results on real images; existing real SR datasets only have fixed magnifications.
Approach
Capture fine-grained, continuously zoomed LR-HR image pairs with a motor-driven lens rig and two-stage SIFT alignment, and propose LMI, which mixes features and relative coordinates of 16 neighboring local tokens via an MLP-mixer with a meta-learned spatial mix weight.
Key results
- LMI with EDSR-baseline reaches 28.86 dB PSNR at x2 on the COZ test set, versus 28.72 for LIIF, with 87.9K parameters.
- Models trained on real COZ data outperform bicubic-trained counterparts on the COZ test set (e.g., LMI x4 PSNR 25.22 vs 24.83 dB with BD training).
- In user studies with 20 participants on real phone/camera images, LMI was ranked highest and real-data-trained models received more votes than simulated-data-trained ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on real COZ image pairs instead of bicubic-downsampled data | arbitrary-scale SR models · COZ test ×2, ×3, ×4 | PSNR: 28.35 / 26.23 / 24.83 -> 28.86 / 26.66 / 25.22 | Helped |
| use LMI instead of the strongest listed baseline | COZ arbitrary-scale SR · COZ test, scales ×2 to ×6 | PSNR (dB): 28.72 -> 28.86; 26.61 -> 26.66; 25.17 -> 25.22; 23.26 -> 23.29 | Helped |
| randomly select two different-resolution images to form LR-HR pairs rather than fixing the highest-resolution HR image | LMI · COZ test, scales ×2, ×3, ×4, ×5, ×6 | PSNR (dB): 28.77 / 26.57 / 25.16 / 24.32 / 23.25 -> 28.86 / 26.66 / 25.22 / 24.39 / 23.29 | Mixed |
| add meta-learning instead of directly embedding coordinates (ablation) | LMI MSMM · COZ test ×2, ×3, ×4, ×5, ×6 | PSNR (dB): 28.77 / 26.58 / 25.15 / 24.34 / 23.23 -> 28.86 / 26.66 / 25.22 / 24.39 / 23.29 | Helped |
| add coordinate information to multi-feature mixing (ablation) | LMI MSMM · COZ test ×2, ×3, ×4, ×5, ×6 | PSNR (dB): 28.74 / 26.55 / 25.13 / 24.32 / 23.20 -> 28.86 / 26.66 / 25.22 / 24.39 / 23.29 | Helped |
| add the MSMM module (ablation) | LMI · COZ test ×2, ×3, ×4, ×5, ×6 | PSNR (dB): 28.67 / 26.51 / 25.11 / 24.31 / 23.18 -> 28.86 / 26.66 / 25.22 / 24.39 / 23.29 | Helped |
| add LR RGB embedding to QMM (ablation) | LMI · COZ test ×2, ×3, ×4, ×5, ×6 | PSNR (dB): 28.81 / 26.64 / 25.22 / 24.37 / 23.27 -> 28.86 / 26.66 / 25.22 / 24.39 / 23.29 | Helped |
Surprising findings
- Training with HR images randomly selected from the scene's zoom sequence degraded performance at x2 relative to always using the highest-resolution image, whereas random selection improved x3-x6.
Limitations
- Test scales are limited to fixed scales near x2-x6, and training magnification only spans x1.0-x4.0.
- Training set is smaller than simulated datasets such as DIV2K, and only a single camera was used for capture.
- Evaluation on real images without ground truth relies on user studies and visual comparison.
Experimental setup
Datasets: COZ
Benchmarks: COZ test set
Metrics: PSNR SSIM user preference ranking
Baselines: MetaSR LIIF LTE LINF SRNO LIT
Builds on: LIIF MetaSR LTE MLP-Mixer SIFT
Closest papers by meaning
- Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling (ICLR 2026 · similarity 0.88)
- Camera Lens Super-Resolution (CVPR 2019 · similarity 0.88)
- Zoom to Learn, Learn to Zoom (CVPR 2019 · similarity 0.87)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.86)
- Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder (CVPR 2024 · similarity 0.86)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · 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.