New Release: Ultralytics v8.4.46

Discussion
, ,
1

Ultralytics v8.4.46 is out :rocket:

Summary: Ultralytics v8.4.46 is a focused stability release that improves training robustness, export reliability, and documentation quality across Ultralytics YOLO workflows. The headline fix is an important multi-scale training edge case correction, alongside safer resume behavior, clearer RKNN support, and better export logging. :raising_hands:

Highlights

:wrench: Multi-scale training fix

The biggest change in v8.4.46 is a priority fix for multi-scale training minimum size clamping. Multi-scale random resizing now enforces a safe lower bound of at least one model stride, helping prevent invalid tiny image sizes during training.

This is especially helpful for:

  • small-image training setups
  • aggressive shrinking or augmentation configurations
  • avoiding crashes and instability during training

This update also adds a regression test so this edge case stays fixed going forward. Huge thanks to @glenn-jocher in PR #24394. :white_check_mark:

:stop_button: Safer resume behavior

Resuming from an already completed checkpoint now correctly exits with a “nothing to resume” message instead of running an extra epoch.

That means:

  • less wasted compute
  • more predictable training automation
  • cleaner resume logic in long-running jobs

Thanks to @lmycross in PR #24386. :clap:

:package: Better export experience

A few useful export improvements landed in this release:

  • Export completion logs now point to the exact artifact path, not just the parent folder, thanks to @lakshanthad in PR #24316
  • Export argument behavior was aligned across code and docs, especially for data and fraction in INT8 calibration workflows, thanks to @lakshanthad in PR #24382

These changes should make export pipelines easier to debug and automate. :rocket:

:brain: Clearer RKNN support

RKNN export behavior is now more explicit for Rockchip users. INT8-only targets such as rv1103, rv1106, rv1103b, and rv1106b are now blocked with a helpful error instead of failing later in the process.

Docs were also updated to reflect current behavior more clearly: FP16-supported RKNN targets only.

Thanks to @lakshanthad in PR #24384. :handshake:

Documentation improvements :books:

This release also includes several documentation polish updates:

These updates help keep the docs more consistent and trustworthy for both new and experienced users.

Why this release matters :bullseye:

Overall, v8.4.46 helps make YOLO workflows more dependable by delivering:

  • more robust training with safer multi-scale resizing
  • less wasted time when resuming finished runs
  • cleaner deployment workflows with better export reporting
  • clearer hardware compatibility feedback for RKNN users
  • higher-quality docs with fewer mismatches and warnings

If you’re training or exporting regularly, this is a very worthwhile upgrade.

What’s changed

Try it out :test_tube:

You can upgrade with:

pip install -U ultralytics

If you’re starting a new project, we recommend using Ultralytics YOLO26, our latest stable model family, and managing training and deployment through the Ultralytics Platform, the simplest way to annotate datasets, train, deploy, and monitor YOLO models.

Release links

See the v8.4.46 release page for full release details, and browse the full changelog from v8.4.45 to v8.4.46 for everything included.

Feedback welcome :speech_balloon:

Please give v8.4.46 a try and let us know how it works for your training, export, or deployment workflows. Community feedback helps the whole Ultralytics ecosystem keep improving.