New Release: Ultralytics v8.4.189

:rocket: Ultralytics v8.4.189: More Reliable Exports and Robust Inference

Ultralytics YOLO v8.4.189 is here, bringing deployment fixes for Sony IMX500 and Hailo, broader ONNX export compatibility, clearer dataset errors, and safer inference results handling. No model architecture changes were introduced in this release.

:package: Export and Deployment Improvements

  • More reliable Sony IMX500 exports: Updated converter requirements and aligned dependency versions help reduce setup and export failures. IMX500 export is now explicitly supported on Linux with Python 3.10–3.12. See PR #26883 by @glenn-jocher for details.
  • Fixed Hailo segmentation exports: The segmentation prototype output node is now detected from the exported model, accommodating different model activations. This fix is also included in PR #26883 by @glenn-jocher.
  • Broader ONNX compatibility: Replacing several einsum operations with equivalent matrix multiplications enables affected YOLOE and YOLO-World models to export at lower ONNX opsets. See PR #26881 by @Y-T-G for the changes.
  • Safer TensorFlow-family exports: Existing ONNX Runtime variants, including Windows ML and Vitis AI builds, are recognized to avoid unnecessary runtime replacement. See PR #26870 by @cainiao33 for details.

:hammer_and_wrench: Inference and Dataset Fixes

  • Stable classification results: Saved probabilities remain unchanged across inference calls, even when a backend reuses its output buffer. See PR #26860 by @aswanth-07 for the fix.
  • Safer grayscale handling: Two-dimensional grayscale NumPy images no longer trigger an IndexError in Annotator and save_one_box. See PR #26878 by @Rahulbiradar9 for details.
  • More resilient results utilities: Results.save_crop and verbose handle results without boxes more safely. See PR #26868 by @wizzseen for the changes.
  • Clearer dataset troubleshooting: Unreadable datasets are less likely to be misleadingly reported as having no labels, thanks to PR #26883 by @glenn-jocher. Classification dataset URL scheme handling is also fixed in PR #26879 by @Vaishnavi220506.

:mobile_phone: Documentation and Benchmarks

  • Mobile app guidance: Ultralytics Platform documentation now explains how to open compatible exports in the Ultralytics YOLO app. See PR #26875 by @glenn-jocher for the update.
  • Refreshed mobile benchmarks: LiteRT, QNN, and Core ML performance tables now use Ultralytics 8.4.188 assets. See PR #26874 by @glenn-jocher for the measurements.
  • Clearer optimizer guidance: Documentation explains momentum behavior when using optimizer=auto. See PR #26877 by @mh-masum6875—and welcome to this first-time contributor! :tada:

:high_voltage: Try It and Share Your Feedback

Install this release with:

pip install --upgrade ultralytics==8.4.189

Explore the v8.4.189 release notes for the overview, or review the full changelog to see every change.

How does v8.4.189 work with your models and deployment targets? Share your results, questions, or issues below—your feedback helps improve future releases. :raising_hands: