New Release: Ultralytics v8.4.190

Discussion
, ,
1

:rocket: Ultralytics v8.4.190: Smarter YOLO26 Segmentation Training

Ultralytics YOLO v8.4.190 is here! This release introduces mask-aware label assignment for YOLO26 segmentation training, adds RKNN INT8 export for depth models, and improves CoreML reliability. You’ll also find faster segmentation validation, better batch sizing, and fixes across inference, data handling, and tooling.

:performing_arts: New Features

  • Mask-aware YOLO26 segmentation training: PR #26773 by @Bovey0809 makes positive training matches consider both box and mask overlap. Bounded mask decoding and safer precision help control memory use and avoid numerical overflow. Other segmentation models and inference/export behavior remain unchanged.
  • RKNN INT8 export for depth models: PR #26885 by @Bovey0809 enables INT8 export with hybrid precision that keeps error-prone layers in FP16. For calibration, use paired depth maps covering your deployment scene’s depth range.

:high_voltage: Performance and Training Improvements

  • Faster segmentation validation: PR #26898 by @raimbekovm vectorizes COCO mask RLE encoding, reducing encoding time in tested workloads.
  • OpenVINO CPU prefetching: PR #26897 by @var1914 allows upcoming image batches to load while the current batch runs.
  • More accurate AutoBatch sizing: PR #26894 by @raimbekovm fixes memory accounting so batch selection better reflects GPU memory available to the process.
  • Earlier validation feedback: PR #26891 by @raimbekovm rejects unsupported semantic single_cls configurations before scanning masks.

:mobile_phone: Export and Inference Fixes

  • No duplicate CoreML segmentation detections: PR #26862 by @amanharshx fixes duplicate detections in segmentation exports with NMS.
  • Safer CoreML inference on MPS: PR #26896 by @amanharshx fixes crashes caused by float64 outputs.
  • Fewer dependency disruptions: PR #26863 by @amanharshx avoids unnecessary NumPy downgrades during CoreML inference.
  • Improved oriented-box NMS precision: PR #26823 by @wpyaoxueros fixes FP16 precision issues in embedded OBB NMS.

:card_index_dividers: Data, Usability, and Tooling

  • Correct handling of unlabeled classification images: PR #26865 by @reubenjrouse makes NDJSON conversion skip unlabeled images instead of assigning them to class 0.
  • More robust FastSAM prompting: PR #26830 by @aswanth-07 handles text prompts when no masks qualify.
  • Floating-point region coordinates: PR #26895 by @Rahulbiradar9 fixes a SolutionAnnotator crash when drawing regions with floating-point coordinates.
  • OCR-ready CPU runner: PR #26893 by @glenn-jocher adds Tesseract OCR with English data for scanned-PDF workflows.
  • CUDA test configuration fix: PR #26900 by @fcakyon parses YOLO_AUTODEVICE_FRACTION_FREE as a float.

:books: Documentation Updates

  • Clearer Ultralytics Platform guidance: PR #26888 by @raimbekovm clarifies partial dataset and project cloning, prediction examples, providers, and the mobile import link.
  • Better VLM discovery: PR #26884 by @glenn-jocher connects VLM documentation to the VLM Leaderboard and Platform model pages.

:raising_hands: Thank You, Contributors!

Thanks to everyone who helped ship this release! A special welcome to @var1914, @reubenjrouse, and @wpyaoxueros for their first contributions.

:hammer_and_wrench: Try It and Share Your Feedback

Upgrade to the latest release:

pip install -U ultralytics

Explore the v8.4.190 release notes for release details, or browse the full changelog to review all changes.

How does v8.4.190 work on your models and hardware? Share your results, questions, and any issues below—especially if you’re training YOLO26 segmentation models or deploying with CoreML, RKNN, or OpenVINO. :rocket: