New Release: Ultralytics v8.4.174

Discussion
, ,
1

:rocket: Ultralytics v8.4.174: Better YOLO26 AutoBatch Profiling and Faster Prediction Logs

Ultralytics YOLO v8.4.174 is here! This release fixes backward-pass profiling for YOLO26 AutoBatch and speeds up prediction summaries, alongside improvements to exports, tracking, inference, and dataset handling.

No model architecture, checkpoint, or default changes are introduced. Here’s what’s improved. :backhand_index_pointing_down:

:glowing_star: Main Highlight: More Accurate AutoBatch Profiling

  • YOLO26 backward profiling and faster prediction logs: PR #26547 by @aswanth-07 makes the profiler find tensors inside nested model outputs, enabling backward profiling for YOLO26 detection heads and other task heads. Prediction summaries also skip unnecessary tensor conversions.

What this means: AutoBatch can measure YOLO26 training memory more accurately, so automatically selected batch sizes may change. Fixed-batch training and prediction are unaffected by the profiling fix.

:package: Export and Inference Improvements

  • Larger dynamic CoreML batches: PR #26551 by @amanharshx processes inputs in chunks when prediction batches exceed the exported batch size.
  • Reusable Hailo calibration data: PR #26544 by @nivosco materializes calibration images before INT8 optimization, supporting methods that revisit the data.
  • More relevant INT8 warnings: PR #26528 by @amanharshx fixes misleading calibration warnings for CoreML and MNN.
  • Predictor and tracker refresh: PR #26541 by @raimbekovm rebuilds predictors and persisted trackers when relevant setup arguments change.
  • Reliable model-format detection: PR #26549 by @raimbekovm uses the terminal file suffix to handle compound filenames correctly.
  • More flexible inputs and class names: PR #26536 by @glenn-jocher preserves non-synset class names and allows YouTube streams below 1080p.

:hammer_and_wrench: Detection and Tracking Fixes

  • Zero-threshold Fast-NMS: PR #26522 by @aswanth-07 keeps OBB detections when the IoU threshold is zero.
  • Correct unconfirmed-track handling: The initial fix in PR #26525 by @wizzseen was reverted through PR #26532 by @glenn-jocher. The refined fix in PR #26543 by @wizzseen hides unconfirmed tracks when appropriate without clearing first-frame detections.
  • More accurate coordinate scaling: PR #26546 by @MaverickTopG uses rounded letterbox dimensions for per-axis gains in scale_boxes and scale_coords.

:card_index_dividers: Dataset and Image Handling

:books: Documentation and Packaging

  • Per-object depth guidance: PR #26521 by @Vaishnavi220506 documents depth estimates with instance segmentation.
  • AMD depth benchmarks: PR #26518 by @lakshanthad adds YOLO26 Depth benchmarks to the AMD documentation.
  • Updated Ultralytics Platform workflows: PR #26538 by @raimbekovm updates guidance for image search, live-camera inference, and class-name matching.
  • Docker improvements: PR #26533 by @glenn-jocher enables Python bytecode caching and standardizes installation with uv.
  • Compatible Conda resolution: PR #26542 by @glenn-jocher resolves Python and Ultralytics together in CI and documentation.
  • Aligned AMD GPU CI dependencies: PR #26550 by @onuralpszr pins ROCm 10.0.0 PyTorch wheels that match the MIGraphX packages.

:raising_hands: Community Thanks

Thank you to the YOLO community and Ultralytics team for these improvements! A special welcome to @MaverickTopG, whose first contribution improves coordinate scaling.

:rocket: Try v8.4.174

Install this release with:

pip install --upgrade ultralytics==8.4.174

Explore the v8.4.174 release notes for release details, or review the full changelog to see every change.

Give it a try and share your feedback below! We’d especially love to hear how AutoBatch, exports, and tracking behave in your workflows.