Research map / Image generation / Super-resolution
Block-Based Multi-Scale Image Rescaling
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 BBMR, a block-based multi-scale image rescaling framework that assigns per-block downscaling factors (x2/x4/x8) at constant overall rate, paired with a JointSR upscaler with a deblocking branch for 2K/4K images.
Problem
Image rescaling for high-resolution (2K and above) images treats the image with a single overall scaling rate, ignoring that information and super-resolution difficulty are unevenly distributed across regions.
Approach
Split the HR image into equal 128x128 sub-blocks and allocate each block a scaling factor (x2, x4 or x8) by ranking PSNR gains from a lightweight SR model, keeping the average scaling rate fixed at x4; a JointSR network upscales mixed-resolution blocks at the feature level and uses a deblock branch to remove block seams.
Key results
- BBMR-OmniSR reaches 34.30 dB PSNR on Test4K vs 33.15 dB for OmniSR with bicubic downscaling at the same overall x4 rate.
- With Down-Net downscaling, BBMR-OmniSR gains up to about 1.97 dB PSNR on Test4K (41.36 vs. 39.60 on Test2K is the reported Down-Net pair), averaging about 1.5 dB improvement.
- Upscaling-module FLOPs increase only about 3% (e.g., 327.87G to 337.22G on Test4K), and a one-OSAG lightweight proxy cuts downscaling-module FLOPs to about 26% at a PSNR cost of 0.12 to 0.21 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply block-based multi-scale rescaling | OmniSR with BICUBIC downscaling · Test2K | PSNR: 36.50 -> 37.78 | Helped |
| Apply block-based multi-scale rescaling | OmniSR with BICUBIC downscaling · Test4K | PSNR: 33.15 -> 34.30 | Helped |
| Apply block-based multi-scale rescaling | OmniSR with Down-Net downscaling · Test2K and Test4K | PSNR: 39.60 -> 41.36; 36.75 -> 38.49 | Helped |
| Apply block-based multi-scale rescaling | CRAFT with BICUBIC downscaling · Test2K and Test4K | PSNR: 36.32 -> 37.53; 32.99 -> 34.05 | Helped |
| Apply block-based multi-scale rescaling | CRAFT with Down-Net downscaling · Test2K and Test4K | PSNR: 39.46 -> 41.43; 36.64 -> 38.46 | Helped |
| Divide the image into blocks using one shared scaling rate (ablation) | OmniSR image rescaling · Test4K | PSNR: 33.15 -> 33.08 | Hurt |
| Add block-based multi-scale downscaling (ablation) | OmniSR image rescaling · Test4K | PSNR: 33.08 -> 34.26 | Helped |
| Add JointSR without the deblock branch (ablation) | BBMR with OmniSR · Test4K | PSNR: 34.26 -> 34.32 | Helped |
| Add the deblock branch to JointSR (ablation) | BBMR with OmniSR · Test4K | PSNR: 34.32 -> 34.30 | Hurt |
| Add the deblock branch to JointSR (ablation) | BBMR with OmniSR · Test4K | Blocking artifacts | Helped |
| Use a lightweight SR model for scaling-rate allocation (ablation) | BBMR-OmniSR · Test2K | PSNR: 37.78 -> 37.57 | Hurt |
| Use a lightweight SR model for scaling-rate allocation (ablation) | BBMR-OmniSR · Test4K | PSNR: 34.30 -> 34.11 | Hurt |
Surprising findings
- The deblock branch slightly lowers PSNR (0.02 dB on Test4K) while visibly removing block artifacts.
- A lightweight SR proxy (one OSAG module) can assist scaling-rate allocation with only a small PSNR loss.
Limitations
- Gains reported only for a fixed 128x128 block size and x2/x4/x8 factor set on the authors' own Test2K/Test4K splits.
- Scope limited to storage/transmission use cases where LR images are not viewed directly.
Experimental setup
Datasets: DIV2K Flickr2K DIV8K
Benchmarks: Test2K (from DIV8K) Test4K (from DIV8K)
Metrics: PSNR NIQE PI
Baselines: BICUBIC downscaling with OmniSR BICUBIC downscaling with CRAFT Down-Net with OmniSR Down-Net with CRAFT
Builds on: OmniSR CRAFT Down-Net
Closest papers by meaning
- Plug-and-Play Tri-Branch Invertible Block for Image Rescaling (AAAI 2025 · similarity 0.89)
- Deep Generative Model based Rate-Distortion for Image Downscaling Assessment (CVPR 2024 · similarity 0.86)
- Sequence Matters: Harnessing Video Models in 3D Super-Resolution (AAAI 2025 · similarity 0.86)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.86)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.86)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · 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.