New Release: Ultralytics v8.3.135

Discussion
,
1

:rocket: Ultralytics v8.3.135 Release Announcement

We’re excited to announce the release of Ultralytics v8.3.135! This update brings speed, reliability, and usability improvements that make working with YOLO models even more seamless—whether you’re training, predicting, or deploying at scale.

:glowing_star: Summary

This release improves model export reliability, speeds up font handling, enhances pose documentation, and streamlines video processing for a smoother user experience. :rocket:

:new_button: New Features & Major Improvements

:shield: Export Stability

  • TensorFlow Export Compatibility: ONNX version is now pinned to >=1.12.0 and <1.18.0 during TensorFlow SavedModel exports, reducing export errors and boosting reliability.
    — PR #20640 by @RizwanMunawar

:high_voltage: Faster Predictions

  • Font Checking Optimization: The result of font checking is now cached, making repeated predictions (yolo predict) much faster.
    — PR #20644 by @jonashaag

:clapper_board: Smarter Video Processing

  • Improved Quit Handling: The “quit” option for video processing is now only shown when results are displayed, which keeps things efficient for headless or automated runs.
    — PR #20643 by @RizwanMunawar

:memo: Clearer Pose Metrics

  • Documentation Upgrade: Pose task documentation now explains new evaluation metrics, making it easier to interpret pose detection results.
    — PR #20659 by @RizwanMunawar

:robot: YOLOE Prompt Handling

  • Reference Prompt Logic: YOLOE now intelligently reuses prompts when you provide a reference image—streamlining predictions across multiple images.
    — PR #20661 by @Y-T-G

:bullseye: Purpose & Impact

  • Smoother Model Exports: Exporting models (e.g., to TensorFlow/TFLite) is more reliable and user-friendly.
  • Faster Workflows: Batch and repeated predictions are now even snappier.
  • Better Automation: Video processing is improved for cases where no display is present.
  • Easier Result Interpretation: Pose documentation updates help you understand model outputs better.
  • Simplified Multi-image Work: YOLOE’s smarter prompt usage improves efficiency when using reference images.

:hammer_and_wrench: Other Notable Changes

:clap: Thanks to Our Contributors

Special shout-out to @jonashaag for their first contribution!
Read more from Cache check_font result PR.

And big thanks to the whole YOLO community and Ultralytics team for driving these improvements forward!

:magnifying_glass_tilted_left: Full Changelog & Release Details

:vertical_traffic_light: Try It Out & Share Feedback!

We encourage everyone to update to v8.3.135, put the new features and fixes to the test, and let us know your thoughts right here in the forum. Your feedback drives future releases—don’t hesitate to share your ideas, bug reports, and suggestions.

Thank you for being a part of the Ultralytics community!

Happy detecting,
The Ultralytics Team