📸

Image Upscaling - LatticeNet

🛠

Parameter-Efficient Image Upscaling (50% Reduction)

Lightweight super-resolution model using lattice blocks with attention-weighted butterfly structures. Designed for resource-constrained environments (mobile, edge devices).

What it does

LatticeNet performs single image super-resolution by stacking lattice blocks that combine two residual blocks via attention-weighted butterfly structures inspired by lattice filter banks. Upscales low-resolution images (2×, 3×, 4×) while maintaining state-of-the-art accuracy with significantly fewer parameters through efficient feature combination and backward feature fusion.

Problem it solves

  • Lightweight SR for resource-constrained environments (mobile, edge, IoT)
  • Parameter efficiency (50% reduction) without sacrificing accuracy
  • Efficient residual block combination via attention mechanisms
  • Backward feature fusion for hierarchical refinement
  • Fast inference on limited computational resources

Input/Output

  • Input: Low-resolution RGB or grayscale image
    • image
  • Output: Upscaled high-resolution image
    • Scale factors: 2×, 3×, or 4× upscaling
    • Resolution: Scale factor × input resolution
    • Example: 64×64 LR → 256×256 HR (4×)
    • Quality: State-of-the-art super-resolution quality with reduced parameters
      • image

Performance

  • Accuracy: State-of-the-art SR quality with 50% parameter reduction
  • Parameter Efficiency:
    • 50% parameter reduction vs. standard SR models
    • Maintains competitive accuracy despite reduced parameters
    • Suitable for mobile/edge deployment
    • Faster training due to smaller model size
  • Metrics:
    • Competitive performance on SR benchmarks (Set5, Set14, BSD100, Urban100)
    • Qualitative visual quality restoration at reduced computational cost

Technical Details

Architecture
Lattice block with butterfly structures
Core Innovation
Lattice blocks combine two residual blocks via an attention-weighted mechanism, reducing parameters by 50% compared to sequential residual blocks.
Core Components
• Lattice Block (LB): Contains two butterfly structures that adaptively combine residual block outputs • Butterfly Structures: Inspired by lattice filter banks, learn efficient feature path routing • Attention Mechanism: Learns optimal coefficients for weighting different feature pathways • Backward Feature Fusion: Hierarchical refinement through multi-scale context extraction
Efficiency Advantages
• 50% parameter reduction • Faster training and inference • Lower memory usage • Suitable for embedded devices and mobile GPUs
Training
• Dataset: DIV2K (standard super-resolution dataset) • Degradation: Bicubic downsampling • Loss: L1, perceptual, or adversarial loss • Scale factors: 2×, 3×, 4× (separate models per scale)
Requirements
• PyTorch (v1.4+) • Supports variable input resolutions

Compliance & Provenance

Provider
Open-source
Provider type
Specialized
License
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