Ultralytics v8.4.215: Leaner Segmentation, More Reliable Workflows
Ultralytics v8.4.215 is here, bringing lower-memory segmentation workflows, more reliable annotations and video processing, and stronger tracking and checkpoint provenance.
This release preserves existing outputs and prediction defaults—no model architecture changes are introduced.
Highlight: Lower-Memory Segmentation
Segmentation workflows can now plot, save, display, move, convert, and prompt masks stored within their bounding boxes, without first building a full image-sized mask stack.
This reduces memory and processing overhead for workflows that only need box-local masks, especially with large images or many detections. Existing behavior remains compatible: requesting masks.data still materializes the dense masks.
Thanks to @glenn-jocher for implementing this improvement in PR #27216 while preserving existing outputs.
Bug Fixes and Reliability
- Correct automatic annotation labels: Class labels now stay aligned with detections when SAM filters out masks. Thanks to @aswanth-07 for the fix in PR #27214 that keeps annotations consistent.
- More complete video predictions: Video loading now reads until decoding ends instead of relying on potentially inaccurate frame-count metadata. Thanks to @Vaishnavi220506 for addressing skipped-frame risks in PR #27215 with this loading update.
- Safer tracking fallback: When ECC camera-motion estimation fails, tracking retains the identity transform rather than potentially using a partially modified estimate. Thanks to @AhmetMert01 for strengthening this fallback in PR #27219 for more dependable tracking.
- Accurate checkpoint provenance: Checkpoints record the Git commit present at run launch, even if the checkout changes later. Thanks to @fcakyon for improving reproducibility in PR #27213 through accurate launch metadata.
Brain Tumor Dataset Update
The updated dataset contains 1,115 brain MRI images, a single tumor class, and a patient-based validation split.
Important: This dataset is for research and education, not clinical diagnosis. It contains no tumor-free examples, which remains an important limitation.
Thanks to @raimbekovm for updating the dataset and its guidance in PR #27218 to support more appropriate use.
Documentation and CI Improvements
- Clearer settings documentation: @miles-deans-ultralytics gave the settings page a distinct title in PR #26595 for easier identification.
- More concise page titles: @miles-deans-ultralytics shortened front-matter titles across 25 documentation pages in PR #26316 to improve clarity.
- More resilient CI: @fcakyon added a single retry for failed asset-cache restores in PR #27220 to help CI recover from transient failures.
Try It and Share Your Feedback
Install this release with:
pip install --upgrade ultralytics==8.4.215
Explore the full changelog for all changes, then give the update a spin.
How does it perform on your workloads? Share your segmentation memory results, video-processing experiences, or any issues in the replies. Thanks to everyone who contributed! ![]()