Skip to content

v0.1.15 Release Notes

Latest

Choose a tag to compare

@chaofengc chaofengc released this 18 Mar 09:06
· 15 commits to main since this release

πŸ“’ Major Changes

🚩 New Metrics

  • Add afine metric proposed in AFINE for No-Reference image quality assessment. Thanks to Du CHEN for their great work πŸ€—, refer to their official paper for more details.
  • Add sfid, a commonly used spatial FID metric for evaluating generative models.
  • Add maclip metric proposed in MACLIP, introducing magnitude in IQA. Thanks to zhicheng for their contribution πŸ€—.
  • Add dmm (paper), a Full-Reference IQA method using SVD-based debiased mapping to mitigate perception bias. Thanks to baoliang for their great work πŸ€—.

πŸ› οΈ Enhancements

  • Bundled the full Q-Align (qalign) architecture (Llama2 + mPLUG-Owl2 model implementations) directly into the package, reducing external dependencies and improving reliability. 9b94a86
  • Improve qalign compatibility with latest transformers. 83e54b7
  • Fixed compare2score inference results with a dedicated mPLUG-Owl2 model variant. 79b3434
  • Enhanced QAlign class with improved image processing and more robust error handling. 317888e
  • Add score direction information (lower_better / higher_better) to all model cards. 4fd66f3

πŸ› Bug Fixes

  • Fix brisque metric when loading from local model weights. 6edfb07
  • Fix GFIQA dataset path (GFIQA β†’ GFIQA-20k). 39c89a6
  • Fix spatial features used in sfid. 9b040e4
  • Remove random permutation in inception_score to keep consistent and reproducible results. 812fc93
  • Fix accelerate version constraint in requirements for compatibility. fe508fe

πŸ“¦ Build System

  • Migrate from setup.py + setup.cfg + MANIFEST.in to modern Python packaging with pyproject.toml and setuptools-scm for dynamic versioning. 2f246e5
  • Add uv installation option for significantly faster dependency resolution:
    pip install uv
    uv pip install pyiqa

πŸ”§ Code Quality

  • Remove unused icecream debug imports from multiple modules. b8e336c
  • Remove class_mapping.json and simplify architecture registration in __init__.py. 317888e
  • Clean up unused code and comments in mPLUG-Owl2 model classes. 317888e
  • Improve pyiqa command line argument descriptions for better usability. 317888e

πŸ‘₯ Contributors

Thank you to the contributors who made this release possible:

  • 🌟 Add afine metric by @ChrisDud0257
  • 🌟 Fix tensor return in afine by @rockerBOO
  • 🌟 Add maclip metric and dmm (#287) by @zhix000
  • 🌟 Add score direction info to model cards by @0Ky

πŸ“ For complete details, see the Full Changelog