New Release: Ultralytics v8.4.135

Discussion
, ,
1

:rocket: Ultralytics v8.4.135 Released

Ultralytics v8.4.135 improves detection reliability by adapting max_det to dataset object counts and standardizing fraction behavior across training and validation.

:glowing_star: New Features

Smarter max_det Selection

For detection, segmentation, pose, and OBB tasks, Ultralytics YOLO now inspects the largest number of labeled objects in a single image.

  • Automatically increases the default max_det when the dataset requires a higher limit.
  • Preserves user-defined max_det values.
  • Warns when a custom limit could restrict validation recall.
  • Propagates the resolved value to native end-to-end model heads before validation, improving consistency for NMS-free models.

This helps prevent crowded images from being silently truncated during validation.

:chart_increasing: Improvements

Clearer Object-Count Warnings

When an image contains more labeled objects than max_det permits, YOLO now explains that:

  • The limit may cap recall and produce misleading validation metrics.
  • Increasing max_det can increase validation and post-processing cost.
  • Model and export-format capacities may still impose prediction limits.

Consistent fraction Behavior

Dataset fraction handling is now more predictable:

  • fraction=1 and fraction=1.0 both use the full dataset.
  • Integers greater than 1 continue to represent an image count.
  • fraction=0 and fraction=0.0 can skip an optional test split.
  • Training and validation splits must contain at least one image.
  • Boolean values such as fraction=True are rejected instead of being interpreted ambiguously.

This removes differences caused by serializers representing the same value as either an integer or float.

:hammer_and_wrench: Documentation and Validation

Training, export, and cloud-training documentation now reflect the normalized fraction semantics. Expanded tests cover:

  • Configuration validation
  • Dataset conversion
  • Concatenated datasets
  • Training pipelines
  • End-to-end detection behavior

:wrench: Bug Fixes

The improved max_det logic reduces the risk of artificially low recall on datasets containing crowded images, while the normalized fraction boundaries eliminate ambiguous configuration behavior.

:package: Pull Requests

:high_voltage: Try v8.4.135

Upgrade to the latest release:

pip install -U ultralytics

Explore the v8.4.135 release notes or review the full v8.4.134 to v8.4.135 changelog.

Give the release a try and share your feedback, especially if you work with crowded datasets, custom max_det settings, or native end-to-end models. :raising_hands: