IQA-PyTorch v0.1.13 Release Notes

📢 Major Changes

📦 Model Weights Migration

• All weights moved to Hugging Face: Now available at chaofengc/IQA-PyTorch-Weights
  • ⚡ Enables faster and more stable downloads
  • 🤝 Allows community contributions for model weights
  • 🌏 For users in mainland China: Use the Hugging Face mirror with:

    export HF_ENDPOINT=https://hf-mirror.com

⚡ Performance Improvements

• Enhanced Metric Efficiency: All metrics now process 1080×800 images with:
  • ⏱️ Sub-second inference time on NVIDIA V100 GPU
  • 💾 Memory efficient: Most metrics require <6GB GPU memory. Exceptions: qalign and qalign_8bit
  • 📊 Detailed results available in efficiency benchmark

📚 Development Updates

• Dataset Organization: Split information consolidated into meta CSV files
  • 🔍 Available at chaofengc/IQA-PyTorch-Datasets-metainfo

🐛 Bug Fixes

• 🔧 Fixed weight loading functionality (5bc4df8)
• 🔄 Removed imgaug for NumPy 2.x compatibility (336831)
• 🛠️ Resolved LIQE num_patch issue (5a3b3e3) - Thanks @avermilov

✨ New Features

📊 New Metrics

• Add perceptual color difference metric msswd proposed in MS-SWD (ECCV2024). (3835ad9)
• Added lpips+ and lpips-vgg+ (b57dede)
• Implemented piqe metric (5b5afb3)
• Added MATLAB-compatible niqe_matlab and brisque_matlab (065c2b7)
• Introduced qalign_8bit and qalign_4bit (e225d8f)

🛠️ Enhanced Functionality

• Added approximate score ranges for metrics (cbba398)
• Implemented optional input validation (0aa6f71)

⚡ Performance & Usability Improvements

🚀 Speed Optimizations

• 📥 Faster loading through simplified imports and lazy loading with class mapper cache
• ⚡ 5x GPU performance boost for nrqm and pi metrics
• 🔋 Enhanced efficiency for ilniqe and piqe through matrix operations

💻 Command Line Interface

• 🔧 Improved CLI syntax:

  pyiqa [metric_name(s)] -t [image_path or dir] -r [image_path or dir] --device cuda --verbose

📖 Documentation

• 📚 Comprehensive documentation updates and improvements

👥 Contributors

Thank you to our new contributors for their valuable input:

• 🌟 @jaycecd: Score range improvements (#180)
• 🌟 @Luciennnnnnn: Cache strategy enhancement (#192)
• 🌟 @RistoAle97: Typo fixes (#197)

📝 For complete details, see the Full Changelog