v0.5.4

• Bug fix: Caching was causing a huge mem-leak

v0.5.3

• Added: ColorVideoVDP can now process batches of images or video (in a PyTorch tensor)
• Bug fix: small discrepancy in the predictions depending on how many frames are processed at once (gpu memory)

v0.5.0

• v0.5.0 (20/June/2025)
  • Added two new metrics: cvvdp-ml-saliency and cvvdp-ml-transformer. See metrics.md for more info.
  • Fixed heat map visualization when the test image has very small dynamic range (thanks to Yuta).

Full Changelog: v0.4.2...v0.5.0

v0.4.2

Changes with respect to v0.4.1:

• Files are now sorted after the wildcard expansion
• Updated PU21 encoding parameters so that they are in sync with those in https://github.com/gfxdisp/pu21/
• Added: --dump-channels for generating videos with intermediate processing stages (debugging and visualization)
• Added: Support for HLG EOTF (e.g. iPhone HDR video) - thanks to Cosmin Stejerean
• Added: Processing of videos stored as image frames, described using the C-notation frame_%04d.png. New arguments: '--fps' and '--frames'
• Fixed: A better memory model for estimating how many frames can be processed at once on a GPU. Added '--gpu-mem' argument.
• Added: 'exposure' field in a display model JSON file.
• Added: --result argument to store results in a CSV file.
• Fixed: Added examples to README.md and improved documentation.
• Added: ColorVideoVDP logo.