Clean Up Those Blocky, Fuzzy JPEGs & Restore Natural Quality
Removes visible block artifacts, color banding, and fuzzy edges from JPEG-compressed photos. That blocky pattern you see in heavily compressed images? Gone. Those weird color transitions in smooth areas? Fixed. Transform degraded JPEGs back to sharp, natural-looking photos.
What it does
JPEG Artiface Removal (SwinIR) removes compression artifacts from JPEG photos and makes them look sharp and natural again. When you save a photo as JPEG with low quality, the file size shrinks but the image gets weird-looking artifacts: blocky patterns, fuzzy edges, color smudging. SwinIR fixes this. It analyzes the artifacts and reconstructs the image to look like it was never compressed. Works on JPEGs at any quality level (QF 10-40), and the lower the original quality, the more improvement you'll see.
Problem it solves
- Blocky/mosaic appearance – Visible square patterns across the image from JPEG compression
- Color banding – Smooth areas show weird color transitions instead of gradual gradients
- Fuzzy/blurry edges – Details and edges look soft or smudged instead of sharp
- Color smudging – Reds and blues look muddy or distorted instead of pure colors
- Old/archived JPEGs – Photos saved with low quality years ago that look terrible
- Downloaded images – Images compressed for web that need restoration
- Mobile/social media photos – Aggressively compressed for smaller file size
- Screenshot quality – Screen captures saved as highly compressed JPEG
Input/Output
- Input: Any JPEG photo or image
- Resolution: Any size (works on 480×480 up to 4K)
- Content: Color or grayscale photos, documents, screenshots
- Compression level: QF 10 (heavily compressed) to QF 40+ (lightly compressed)
- Output: A cleaned-up version of the same photo
- Resolution: Same as input (no upscaling, just artifact removal)
- Quality: Blocks removed, colors natural, edges sharp again
Different Compression Levels:
Heavy Compression (QF 10) | Medium Compression (QF 20–30) | Light Compression (QF 40) |
Most artifacts | Normal JPEG quality | Minimal artifacts |
Performance
- Real-World Quality by Scenario:
- Heavy compression (QF 10): Dramatic improvement, blocks completely removed
- Standard web JPEG (QF 20-30): Clear improvement, artifacts mostly gone
- Light compression (QF 40+): Subtle improvement, already decent quality
Technical Details
- Architecture: SwinIR-M (Swin Transformer for Image Restoration)
- Key Features: Window size 7×7 aligns with JPEG's 8×8 block structure for effective artifact removal.
- Core Components:
- Backbone: Swin Transformer with 6 residual blocks (RSTB)
- Hidden dimension: 180, 6 attention heads
- Parameters: ~18.8M
- Feature Extraction: Shallow + deep extraction with residual connections
- Image Reconstruction: 1×1 convolution with skip connection
- Training:
- Dataset: DIV2K, Flickr2K, BSD500, WED (8,594 images total)
- Training patches: 126×126
- Loss: L1 (MAE) + perceptual loss
- Optimizer: Adam (lr=2e-3)
- Batch size: 128
- Epochs: 500k iterations (~3-4 days on 8× RTX2080 Ti)
- Augmentation: Random crop, flip, rotation
- JPEG Degradation Model:
- Input: JPEG-compressed image (quality factor 10/20/30/40)
- Output: Restored clean image
Compliance & Provenance
Provider | Open-source |
Provider type | Specialized |
License | Apache 2.0 |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- GitHub: https://github.com/JingyunLiang/SwinIR (Official implementation with pretrained checkpoint downloads)
- Paper: Liang et al., "SwinIR: Image Restoration Using Swin Transformer" (ICCV 2022, arXiv:2108.10257) https://arxiv.org/abs/2108.10257
- License: apache-2.0