Ultralytics v8.4.213: Leaner Segmentation, More Reliable Predictions
Ultralytics v8.4.213 makes large segmentation predictions more memory- and compute-efficient, improves tiled inference and dataset conversion, and adds better Ultralytics Platform training previews and deployment guidance.
This release focuses on practical improvements across your workflow—no model architecture changes are included.
Highlight: More Efficient Segmentation Masks
With PR #27197 by @glenn-jocher, large segmentation mask stacks are now decoded inside each detection box, rather than across the full image.
- Less unnecessary processing: Outlines and exports can use box-local masks directly.
- Dense masks on demand: Accessing
masks.datastill produces the familiar full-size tensor. - Unchanged behavior for smaller stacks: These continue using the existing path.
If you work with large images or many detections, this can reduce mask-processing costs—especially when you only need outlines or exports. ![]()
Inference, Dataset, and Visualization Fixes
- More reliable tiled predictions: In PR #27188 by @AhmetMert01, class-aware NMS no longer suppresses tile-seam detections because they overlap an object of a different class.
- Safer COCO conversion: With PR #27187 by @AhmetMert01, conversion now raises an error when different images map to the same label file, instead of mixing their labels.
- Correct SAM class alignment: Through PR #27193 by @Yannis-A-D,
yolo_bbox2segmentkeeps class IDs aligned with masks that survive confidence filtering. - Accurate confusion-matrix percentages: In PR #27183 by @laodouya, Platform results now display correct percentages for models with more than 100 classes.
- Correct red and blue channels: With PR #27196 by @AhmetMert01, color handling is corrected for Annotator PIL images and SAM3 tensor results.
Platform and OCR Improvements
- Initial training previews: With PR #27173 by @Laughing-q, training can upload epoch-0 batch previews to Ultralytics Platform.
- Cleaner project names: In PR #27191 by @glenn-jocher, Platform project naming uses the last folder of a local save directory.
- Consistent OCR runtime setup: Through PR #27189 by @Laughing-q, OCR now shares the ONNX Runtime setup used by Ultralytics YOLO models.
Clearer Documentation
- Expanded deployment guidance: In PR #27190 by @raimbekovm, the ONNX and TensorRT integration pages were reorganized with output formats, deployment details, and refreshed benchmarks.
- Broader workflow coverage: With PR #27194 by @pderrenger, documentation now covers per-class validation exports and updates guidance for Gemini 4 Hydrogen, OCR languages, and Platform search and import.
Welcome and Thanks
A special welcome to @Yannis-A-D, whose first contribution in PR #27193 fixes SAM mask/class alignment. Thanks to everyone who contributed!
Try v8.4.213
Install this release with:
pip install --upgrade ultralytics==8.4.213
Explore the v8.4.213 release notes for 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 the segmentation improvements perform on your large-image and high-detection workloads. ![]()