New Release: Ultralytics v8.4.51

Discussion
, ,
1

Ultralytics v8.4.51 is out :tada:

Quick summary: Ultralytics v8.4.51 is a quality-focused release centered on better training traceability, stronger internal augmentation foundations, and clearer deployment and documentation updates. The headline improvement is that training metadata now includes the Git commit message, making experiments easier to track, reproduce, and audit :receipt:

If you’re training, exporting, or deploying with Ultralytics YOLO, this release should make your workflows a bit more transparent and dependable.

Highlights

:receipt: Better training traceability

The biggest update in v8.4.51 is that the current Git commit subject is now captured in training metadata.

This includes:

  • saved checkpoints as git.message
  • Ultralytics Platform training environment metadata as gitCommitMessage
  • Git utilities via GitRepo.message

This makes it much easier to answer: which exact code change produced this model? Huge help for reproducibility, debugging, and team collaboration.

Implemented in PR #24505 by @glenn-jocher.

:hammer_and_wrench: Unified augmentation pipeline refactor

A major internal refactor introduced a more consistent transform system around BaseTransform.

This update helps standardize how transforms are applied across:

  • images
  • instances
  • semantic masks

It also reorganizes augmentations such as:

  • Mosaic
  • MixUp
  • CutMix
  • CopyPaste
  • RandomPerspective
  • RandomFlip
  • LetterBox

While mostly internal, this is an important foundation for future augmentation improvements and maintainability.

Implemented in PR #24434 by @Laughing-q.

:rocket: OpenVINO docs now feature YOLO26 benchmarks

The OpenVINO documentation has been refreshed to show YOLO26 benchmark results instead of older YOLO11 results, giving users a more current view of deployment performance across newer Intel CPUs, GPUs, and NPUs.

Updated in PR #24285 by @lakshanthad.

Improvements

:wrench: More robust Git metadata handling

Git repository parsing was improved to better support metadata extraction from Git internals, including shared and worktree-style layouts. This helps version information get captured more reliably during training.

Included as part of PR #24505 by @glenn-jocher.

:outbox_tray: Expanded DeepX export documentation

DeepX was added to the export formats table, and export argument documentation was cleaned up for clarity, including output folder behavior.

Implemented in PR #24488 by @lakshanthad.

:high_voltage: RT-DETR inference tuning guidance

Docs now explain that reducing query count can speed up RT-DETR inference, giving users a practical latency/accuracy tradeoff knob.

Added in PR #24466 by @artest08.

:warning: YOLOE export behavior clarified

The docs now clearly warn that exported YOLOE models are static. After export, prompt-based class configuration is baked into the model and cannot be changed later.

Added in PR #24496 by @raimbekovm.

:cloud: Ultralytics Platform GPU docs refreshed

The Ultralytics Platform documentation now reflects updated GPU availability, including B300, along with revised access and pricing details.

Updated in PR #24506 by @glenn-jocher.

Bug Fixes and Cleanup

Why this release matters

v8.4.51 is less about flashy new end-user features and more about making everyday workflows smoother:

  • Easier experiment tracking with human-readable Git commit messages
  • Better reproducibility across local and Ultralytics Platform training runs
  • More reliable deployment decisions with updated YOLO26 benchmark and export docs
  • Stronger augmentation foundations for future releases
  • Fewer surprises in export workflows, especially for YOLOE users

Try it out

You can upgrade with:

pip install -U ultralytics

Then explore the release details in the v8.4.51 release page or browse the full changelog.

If you give v8.4.51 a try, we’d love to hear how it works for you. Please share feedback, report regressions, or let us know which improvements you’d like to see next :raising_hands: