New Release: Ultralytics v8.4.153

Discussion
, ,
1

:rocket: Ultralytics v8.4.153 Released

Ultralytics v8.4.153 fixes SAM3 initialization with compile=False, restores reliable standalone validation, preserves dataset overrides when resuming training, improves INT8 and IMX exports, and expands Ultralytics Platform annotation and billing documentation.

Users working with SAM3, standalone validation, resumed training, IMX export, or RKNN INT8 deployment are strongly encouraged to upgrade. :counterclockwise_arrows_button:

:glowing_star: Key Highlights

  • :brain: SAM3 no longer triggers unintended torch.compile behavior when compile=False.
  • :white_check_mark: Semantic segmentation and classification validation are more reliable.
  • :counterclockwise_arrows_button: Explicit data overrides are preserved when resuming training.
  • :gear: INT8 behavior is clearer across ONNX, TensorRT, and RKNN exports.
  • :bullseye: IMX export now correctly preserves conf=0.0.
  • :puzzle_piece: Ultralytics Platform documentation now covers batch annotation, pose smart annotation, billing, referrals, and more.

:hammer_and_wrench: Bug Fixes

SAM3 initialization

PR #26189 by GitHub author @glenn-jocher correctly converts the SAM3 predictor compilation option:

  • False disables compilation.
  • True enables the default compilation mode.
  • Explicit compilation mode strings remain unchanged.

This prevents unintended torch.compile activation and related PyTorch 2.14 Dynamo errors. The fix was validated across all six Platform SAM models using HTTP and WebSocket inference.

Standalone semantic segmentation validation

PR #26176 by GitHub author @cainiao33 adds background-class metadata while constructing polygon-based semantic datasets.

This prevents background pixels from being assigned to the final real class, restoring meaningful metrics such as mIoU and pixel accuracy.

Standalone classification validation

PR #26182 by GitHub author @cainiao33 filters validation samples with class IDs outside the model’s supported class range.

Instead of raising an IndexError, validation now skips incompatible samples and provides clear warnings about mismatched classes.

Training resume dataset overrides

PR #26175 by GitHub author @cainiao33 ensures that train(resume=True, data="...") honors the explicitly provided dataset rather than silently reverting to the checkpoint’s original dataset.

Path-based resume workflows also avoid injecting an unintended task-default dataset.

IMX confidence thresholds

PR #26183 by GitHub author @cainiao33 preserves an explicit conf=0.0 during IMX export instead of replacing it with 0.001.

Tests now cover both conf=0.0 and the default conf=0.25 behavior.

:gear: Export Improvements

RKNN INT8 validation

PR #26184 by GitHub author @raimbekovm correctly limits RKNN INT8 validation to detection models, including INT8-only Rockchip targets.

The accompanying examples now use a compatible detection model and dataset.

INT8 documentation and deployment guidance

PR #26174 by GitHub author @raimbekovm aligns documentation with current export behavior:

  • Quantization-aware training checkpoints can use stored INT8 ranges for ONNX and TensorRT exports without calibration data.
  • TensorRT documentation no longer states that dynamic is automatically enabled for INT8.
  • The Platform Python requirement and Hailo hardware comparison are updated.
  • Huawei Ascend and Google Coral Edge TPU are included in the Hailo comparison.

:puzzle_piece: Ultralytics Platform Updates

Batch annotation and billing

PR #26186 by GitHub author @raimbekovm documents dataset-wide batch annotation, including:

  • Class mapping
  • Progress tracking
  • Stopping annotation runs
  • Dataset version snapshots
  • Batch annotation pricing and billing

Pose smart annotation and Platform guidance

PR #26185 by GitHub author @raimbekovm expands Platform documentation with:

  • YOLO smart annotation for pose datasets
  • SAM annotation support for detection, segmentation, semantic segmentation, and OBB tasks
  • Supported dataset licenses
  • The referrals tab
  • Endpoint uptime charges
  • Python 3.11 requirements
  • Updated Agents availability without the previous Early Access requirement

:books: Documentation Updates

:package: Upgrade Today

Install or upgrade to the latest release with:

pip install -U ultralytics

Review the complete changes in the v8.4.152 to v8.4.153 comparison, or visit the Ultralytics v8.4.153 GitHub release for release details.

Give v8.4.153 a try and share your results, questions, or feedback with the Ultralytics YOLO community. Happy building! :tada: