New Release: Ultralytics v8.4.168

Discussion
, ,
1

:rocket: Ultralytics v8.4.168: More Reliable Data Handling and Inference

Ultralytics YOLO v8.4.168 brings practical fixes for image and semantic-mask processing, prediction outputs, video timing, and fixed-batch exported models—plus clearer training guidance. No model architecture changes are included.

:framed_picture: Data Handling and Prediction Fixes

  • Safer masks, fresher caches, and distinct output files: Palette PNG masks retain their class IDs, disk caches refresh when source images change, and AVIF images with EXIF orientation load upright. Predictions for same-stem files such as bus.jpg and bus.png now receive distinct output names. Fixed-batch exports also accept short batches by padding inputs and removing extra outputs. These fixes come from @cainiao33 in PR #26443.
  • Preserved semantic-mask IDs and video timing: 16-bit semantic masks retain their IDs and flag invalid classes, fractional video frame rates are preserved to help avoid playback drift, and CLI run-name parsing is corrected by @Nicholas022400701 in PR #26440.
  • Working CutMix for semantic masks: Semantic-mask regions can now be pasted even when there are no object instances, fixing skipped augmentation on PNG semantic-mask datasets through @Nikhi00718’s PR #26439.

:gear: Runtime and Testing Improvements

  • Quieter ReID initialization: Re-identification models warm up with an in-memory image, avoiding a missing-source warning and unnecessary asset-image inference thanks to @raimbekovm in PR #26453.
  • Lighter similarity-search tests: SearchApp testing reuses existing test images instead of triggering an additional 32 MB download, as updated by @raimbekovm in PR #26452.

:books: Clearer Training and Deployment Guidance

  • Understand what mask_ratio controls: Documentation now clarifies that it affects training masks—not predicted mask resolution—with the update from @Y-T-G in PR #26441.
  • Set expectations for automatic optimization: Documentation explains that optimizer=auto ignores lr0 and momentum, clarifies deployment rename and slug behavior, and adjusts CLI examples through @raimbekovm’s PR #26447.

:rocket: Try It and Share Your Feedback

Upgrade to the release:

pip install --upgrade ultralytics==8.4.168

Explore the v8.4.168 release notes for the release overview, or review the full changelog to inspect all changes.

Try it with your datasets and inference workflows, and share your results or any issues you encounter in this discussion. :raising_hands: Thanks to the contributors, the YOLO community, and the Ultralytics team for making this release possible!