ultralytics 8.3.216

v8.3.216 - `ultralytics 8.3.216` Faster `Annotator` mask plotting (#22419)

on Python PyPI

🌟 Summary

Ultralytics 8.3.216 speeds up and stabilizes segmentation mask rendering, improves pose keypoint metadata handling, and enhances checkpoint loading compatibility across PyTorch versions. ⚡🖼️🧩

📊 Key Changes

• Faster and more reliable mask plotting in the Annotator and Plotter

  • Refactored mask handling and tensor ops for efficient blending and correct device usage.
  • Proper upscaling and padding removal when retina_masks=False using ops.scale_masks.
  • Safer mask processing in ops.process_mask with broadcasted ratios and named arguments.
  • Early exits and streamlined conversions for improved performance.
  • PR: Faster Annotator mask plotting (#22419), Optimize box downsampling for Segment models (#22417)

• Pose models: keypoint names metadata (kpt_names) support end-to-end

  • Dataset YAMLs updated with human, dog, hand, and tiger keypoint names.
  • Trainer attaches model.kpt_names; exporter saves kpt_names into model metadata.
  • AutoBackend can parse kpt_names for downstream use and exports.
  • PR: Support keypoint names metadata for pose models (#22411)

• More compatible checkpoint loading for SAM/SAM 2

  • Uses torch.load(..., weights_only=False) when supported, with safe fallback for older PyTorch.
  • Reduces unexpected keyword errors and deprecation warnings across environments.
  • PR: Include weights_only parameter in torch.load (#22421)

• Documentation enhancement

  • Added a visual cover to the Model YAML Configuration guide for clarity.
  • PR: Add cover image for model-yaml-config.md (#22414)

🎯 Purpose & Impact

• Faster, cleaner segmentation overlays 🎨

  • Users see smoother, more accurate mask visuals with fewer device mismatches and better handling of non-retina masks.
  • Reduced overhead and improved stability for real-time and notebook visualizations.

• Better pose model usability 🕺

  • Human-readable keypoint names seamlessly flow from datasets → training → exports → inference, making results interpretation and downstream tooling easier.
  • Minimal code to use:

    from ultralytics import YOLO
    model = YOLO("yolo11n-pose.pt")
    print(model.kpt_names)

• More robust model loading 🧠

  • SAM/SAM 2 checkpoints load reliably across PyTorch versions, decreasing friction during setup and deployment.

• Improved learning experience 📚

  • The YAML guide’s cover image helps newcomers and experts quickly grasp model configuration structure.

Version bump: 8.3.215 → 8.3.216. 🚀

What's Changed

• Optimize box downsampling for Segment models by @glenn-jocher in #22417
• Change torch.load to include weights_only parameter by @glenn-jocher in #22421
• Support keypoint names metadata for pose models by @Y-T-G in #22411
• Add cover image for model-yaml-config.md by @RizwanMunawar in #22414
• ultralytics 8.3.216 Faster Annotator mask plotting by @glenn-jocher in #22419

Full Changelog: v8.3.215...v8.3.216