Ultralytics v8.4.177: Consistent SAM Prompts and More Reliable Workflows
Ultralytics YOLO v8.4.177 brings consistent SAM prompting across image directories and video frames, plus fixes for model exports, training resumes, prediction inputs, and experiment tracking. This release focuses on reliability; it does not introduce a new model architecture.
Highlight: SAM Prompts Across Images and Video
Supply SAM prompts once and apply them to every image or frame in your source—not just the first. Prompt coordinates are also scaled without modifying caller-owned NumPy arrays or tensors, supporting more predictable Python, streaming, and CLI workflows.
This fix comes from @8rulerstar through PR #26651, which makes prompting consistent across a source.
Export and Prediction Fixes
- Intel NPU exports: PR #26656 by @onuralpszr pins dynamic OpenVINO exports to the requested image size on Intel NPUs, which require fixed input dimensions.
- ONNX GPU inference: PR #26657 by @aswanth-07 adds a compatible execution fallback when output shapes contain symbolic dimensions.
- Image orientation: PR #26650 by @Nikhi00718 applies EXIF orientation to PIL prediction inputs.
- Mixed source lists: PR #26660 by @raimbekovm restricts downloads to actual HTTP(S) URLs, avoiding confusion with local paths beginning with
http.
Training and Experiment Tracking
- Short-run updates: PR #26663 by @raimbekovm ensures runs shorter than the accumulation interval still take an optimizer step.
- Safer RMSprop resumes: PR #26652 by @Vaishnavi220506 preserves small second-moment values in checkpoints, helping prevent unexpectedly large updates after resuming.
- W&B reliability: PR #26654 by @raimbekovm fixes final plot logging, offline resume IDs, project/run naming, and cleanup after interrupted or failed runs.
- Per-run integration settings: PR #26658 by @raimbekovm reads settings for each trainer and keeps event session IDs separate from the seeded random number generator.
Documentation and Workflow Improvements
- Platform exports and contribution testing: PR #26659 by @glenn-jocher clarifies export options and updates testing advice.
- Validation geometry: PR #26653 by @lu123bing documents rectangular validation image geometry.
- Documentation video: PR #26649 by @RizwanMunawar adds a video resource to the docs.
- Centralized CI alerts: PR #26661 by @glenn-jocher moves workflow failure alerts into the GitHub App.
- Codecov verification: PR #26662 by @glenn-jocher restores runner dependencies needed for signature checks.
Community Thanks
Thanks to the YOLO community and Ultralytics team for these improvements! A special welcome to @lu123bing for their first contribution.
Try It and Share Your Feedback
Install this release with:
pip install --upgrade ultralytics==8.4.177
Explore the v8.4.177 release notes for details, or review the full changelog to compare changes.
How does this release work with your SAM, training, or deployment workflows? Share your results and any issues below—reproducible examples help the community investigate!