Ultralytics v8.4.178: More Reliable Platform Links and Stronger Workflows
Ultralytics YOLO v8.4.178 is here! This release improves Platform training links and strengthens validation, tracking, exports, and training integrations. It also introduces a benchmark resource covering 37 real-world detection datasets.
Highlights
- Reliable training links: PR #26699 by @glenn-jocher uses the model URL returned by Ultralytics Platform, reflecting the final owner, project, and model name. If the server provides no URL, no link is printed.
- Broader evaluation coverage: PR #26691 by @Laughing-q adds the Ultralytics Benchmark page with a fixed evaluation protocol across 37 detection datasets, including medical, retail, industrial, and aerial imagery. This is an evaluation resource—not a new model architecture.
- Better YOLO-NAS compatibility: PR #26670 by @raimbekovm fixes loading by model name, prediction, validation, export, and ReID tracking.
Model, Tracking, and Validation Fixes
- Persistent prompts and object IDs: PR #26676 by @raimbekovm keeps FastSAM prompts active across later frames and preserves SAM2 object IDs when masks are blank.
- More consistent oriented-box tracking: PR #26680 by @raimbekovm handles equivalent box representations without disrupting tracking.
- Correct static-export validation: PR #26690 by @Nikhi00718 respects recorded rectangular input shapes.
- Safer SAM mask prompts: PR #26673 by @cainiao33 fixes a crash by downsampling prompts to the encoder’s low-resolution input size.
- Preserved segmentation detail: PR #26667 by @Vaishnavi220506 retains distinct grounding polygons that share the same bounding box.
Export and Hardware Improvements
- TorchScript on Apple MPS: PR #26693 by @raimbekovm fixes loading by converting traced
float64scalars tofloat32. - Dynamic ONNX on MPS: PR #26686 by @raimbekovm keeps dynamic-shape nodes off the CoreML execution provider.
- Full-integer LiteRT inputs: PR #26674 by @aswanth-07 clamps input values before casting.
- Ascend NPU training: PR #26687 by @Eric-Lotus fixes an unsupported reduction in
TaskAlignedAssigner. Welcome to your first contribution!
Platform and Experiment Reporting
- Aligned class labels: PR #26665 by @JaviChulvi sends class indices with Platform validation curves when validation data omits classes.
- Task-specific curves: PR #26678 by @JaviChulvi sends segmentation and pose validation curves from the correct task.
- Correct Comet confusion matrices: PR #26669 by @raimbekovm fixes matrix orientation and labels.
- Cleaner TensorBoard teardown: PR #26682 by @Vaishnavi220506 closes writers after training.
- Default MLflow tracking: PR #26688 by @raimbekovm uses MLflow’s default tracking store when no tracking URI is configured.
Training, Data, and Workflow Reliability
- Safer Docker publishing: PR #26666 by @glenn-jocher prevents overwriting image tags built from newer commits.
- Preserved optimizer precision: PR #26672 by @cainiao33 retains Adam’s
exp_avgmomentum in FP32 during optimizer FP16 conversion. - UTF-8 datasets: PR #26681 by @Nikhi00718 explicitly reads NDJSON datasets with UTF-8 encoding.
- More robust source loading: PR #26671 by @cainiao33 handles UTF-8 BOM source manifests and adds a URL timeout to
autocast_list. - Screenshot inputs: PR #26664 by @Nikhi00718 fixes prediction with monitor and region parameters.
- Configuration directories: PR #26692 by @raimbekovm creates a missing
YOLO_CONFIG_DIRinstead of falling back to/tmp. - Correct ZIP names: PR #26694 by @Vaishnavi220506 preserves full directory names when creating archives.
- Clearer CLI errors: PR #26675 by @glenn-jocher simplifies argument errors and excludes callables from suggestions.
- Cleaner solution output: PR #26697 by @MohammadHijjawi97 skips empty workout lists when printing
SolutionResults.
Documentation Updates
- Annotation model coverage: PR #26668 by @glenn-jocher lists all Platform class-prompted annotation models.
- Image-generation guidance: PR #26683 by @glenn-jocher documents paid image-generation models.
- Platform workflows: PR #26677 by @raimbekovm documents Explore filters, AI-assisted training diagnosis, and API details.
- Improved visuals: PR #26684 by @glenn-jocher adds updated Ultralytics Academy task and guide images.
- Documentation corrections: PR #26685 by @glenn-jocher fixes two figures and retraining-chart alt text.
Try It and Share Your Feedback
Install this release with:
pip install --upgrade ultralytics==8.4.178
Explore the v8.4.178 release notes for the release overview, or review the full changelog for every change.
How does v8.4.178 work with your models, exports, and Platform workflows? Share your results and any issues below. Thanks to the YOLO community and Ultralytics team for keeping these improvements coming! 