New Release: Ultralytics v8.3.237

:rocket: Ultralytics v8.3.237 is live!

TL;DR: v8.3.237 brings full SAM 3 image + video segmentation support (including text + exemplar prompts and tracking), improves export reliability (notably ONNX FP16 on CPU plus cleaner Edge TPU/IMX dependencies), and adds several training/validation + docs quality-of-life upgrades. :brain::movie_camera::gear:

Upgrade any time with pip install -U ultralytics and let us know what you build!


:brain: New Features: SAM 3 (Image + Video) Integration

Ultralytics now supports SAM 3 end-to-end across interactive + semantic (concept) segmentation, plus video tracking:

  • :puzzle_piece: New SAM 3 model builder stack

    • New build_sam3.py and SAM 3 modules (ViT backbone, encoder/decoder, text + geometry encoders, video tracker).
    • The SAM entrypoint now detects sam3.pt and routes builds via build_interactive_sam3.
  • :control_knobs: New SAM 3 predictors & public APIs

    • SAM3Predictor – interactive segmentation
    • SAM3SemanticPredictor – text + exemplar concept segmentation on images
    • SAM3VideoPredictor – video tracking with box prompts
    • SAM3VideoSemanticPredictor – video concept tracking (text + boxes + masklets)
    • Integrated into SAM’s task_map, so SAM("sam3.pt") routes automatically to the correct predictor.
  • :books: Docs upgraded from preview → real usage

    • SAM 3 reference docs added for modules under docs/en/reference/models/sam/sam3/*.
    • docs/en/models/sam-3.md rewritten with concrete examples and important setup notes:
      • SAM 3 weights are not auto-downloaded (you must fetch sam3.pt from the official repo).
      • Text prompts require the BPE vocab file (bpe_simple_vocab_16e6.txt.gz).

For more, see the SAM 3 integration PR by @Laughing-q in PR #22897.


:rocket: Improvements: Export & Device Tooling

:1234: ONNX FP16 export on CPU

ONNX export now supports half=True on CPU, converting weights to FP16 and gracefully warning/falling back if conversion fails (instead of aborting).

Implemented by @glenn-jocher in PR #22927.

:penguin: Edge TPU & IMX export dependency management

Export tooling is now more robust on Debian/Ubuntu-based systems:

  • Edge TPU export now uses centralized check_apt_requirements(["edgetpu-compiler"])
  • IMX export uses check_apt_requirements() for appropriate Java packages
  • check_apt_requirements() now runs apt update with check=True so failures surface clearly

Implemented by @lakshanthad in PR #22925.


:person_lifting_weights: Training & Experiment Quality-of-Life

:counterclockwise_arrows_button: More flexible resume-training overrides

Resuming training now allows overriding more runtime/logging options without restarting:
save_period, workers, cache, patience, time, freeze, val, plots.

Added by @Y-T-G in PR #22903.

:open_book: Data augmentation docs: scale range clarified

Docs updated to clarify scale should be 0.0 - 1.0, avoiding unstable configs.

Updated by @Y-T-G in PR #22907.


:white_check_mark: Validation & Visualization Fixes

:straight_ruler: RT-DETR validation scaling fix

Validation transforms simplified and scaling behavior made explicit/safe via a no-op scale_preds() override.

Fixed by @Y-T-G in PR #22817.

:compass: OBB plotting robustness

OBBValidator.plot_predictions() is now more robust (handles empty predictions cleanly, avoids redundant conversions, uses plot_images() directly).

Fixed by @onuralpszr in PR #22765.


:test_tube: Quick Try

pip install -U ultralytics
from ultralytics import SAM

model = SAM("sam3.pt")  # ensure you've downloaded sam3.pt manually
results = model("image.jpg")

:link: Release + Changelog

Dive into the release details in the v8.3.237 release page, or browse all commits in the full changelog comparison.


:speech_balloon: Feedback welcome!

If you try SAM 3 (especially text prompts or video concept tracking) or the updated export paths, please share:

  • your platform (OS / CPU / GPU)
  • the command or snippet you used
  • any logs or minimal repro steps if something looks off

Your feedback helps keep Ultralytics YOLO and the broader community ecosystem moving fast.