📸

Image Upscaling - SwinIR

🛠

Intelligent Photo Enlargement

Enlarges low-resolution images larger while intelligently recovering lost details. Works like a smart magnifying glass that analyzes image patterns (edges, colors, textures) and reconstructs higher-resolution versions. Perfect for old photos, archives, thumbnails, and preparing images for printing or display.

What it does

SwinIR Classical enlarges small or blurry photographs into larger, sharper, clearer versions. Instead of simple pixel duplication (which looks blocky), it uses AI to understand image patterns and intelligently fill in missing details. The result looks like the photo was originally captured at high resolution. Works on any color photograph—no special preparation needed.

Problem it solves

  • Enlarge small photos – Thumbnails, compressed images, low-res downloads made usable
  • Restore old/faded photos – Archived images, scans, social media downloads restored to sharp condition
  • Prepare for printing – Images enlarged for poster, print, or display without blurriness
  • Enhance video frames – Video screenshots and scanned documents improved for visual clarity
  • Flexible sizing – Choose enlargement factor (2×, 3×, 4×, 8×) based on your need
  • Lost detail recovery – Intelligent reconstruction of textures, edges, and colors
  • No preprocessing needed – Works directly on any photo, any size

Input/Output

  • Input: Any color photograph or image (JPEG, PNG, etc.)
    • image
  • Output: Larger, sharper, clearer version of your image. Size determined by enlargement factor you choose (×2, ×3, ×4, or ×8)
    • image

Performance

  • Quality Metrics (PSNR – how sharp the results are):
Enlargement
Dataset
PSNR
Quality Level
2x
Set5
38.47 dB
Excellent detail
3x
Set5
34.91 dB
Very good detail
4x
Set5
32.92 dB
Good detail (standard)
8x
Set5
27.65 dB
Decent, visible improvement
  • Quality by Content Type:
    • Buildings & patterns: Preserves architectural details and textures well
    • Artwork & illustrations: Maintains line quality and color contrast
    • Faces & people: Good detail recovery, natural appearance
    • Text & documents: Readable after 2-4× enlargement

Technical Details

Architecture
Swin Transformer for Image Super-Resolution (SwinIR)
Key Innovation
Dual-level analysis—local detail focus (8×8 pixel areas for fine textures) combined with global structure understanding (overall shapes and composition) for realistic enlargements.
Core Components
• Shallow Feature Extraction: Initial feature layer • Deep Feature Extraction: Residual Swin Transformer Blocks (RSTB) with shifted window attention and skip connections • Upsampling: Intelligent scaling to final resolution • Image Reconstruction: Final detail enhancement
Training
• Dataset: DIV2K (3,450 high-quality photographs) • Degradation: Bicubic downsampling • Loss: L1 (Charbonnier) + perceptual loss • Optimizer: Adam (learning rate 2e-3, batch size 4) • Training time: 1.6 days on high-end GPUs • Augmentation: Random crop, rotation, flip • Scale factors: 2×, 3×, 4×, 8× (separate models per scale)
Requirements
• PyTorch (v1.8.0+) • Model size: ~11.8 MB • CUDA 10.2+ recommended for GPU inference

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