New Release: Ultralytics v8.4.52

Discussion
, ,
1

:rocket: Ultralytics v8.4.52 is here: native semantic segmentation arrives with YOLO26

:glowing_star: Summary

We’re excited to announce Ultralytics v8.4.52, a feature-packed release led by full native semantic segmentation support across Ultralytics YOLO :tada:

The headline update is simple: YOLO26 now supports semantic segmentation as a first-class task, with new pretrained -sem models, end-to-end training and inference support, export compatibility, broader dataset support, and stronger testing coverage.

If you’ve been waiting to use YOLO for pixel-level scene understanding, this is the release to try. :fire:

:new_button: New Features

:brain: Native semantic segmentation support across YOLO

With v8.4.52, semantic segmentation is now fully integrated into the Ultralytics YOLO workflow alongside detection, instance segmentation, classification, pose, and OBB.

This includes support across:

  • CLI
  • Python API
  • training
  • validation
  • prediction
  • export
  • benchmarks
  • documentation

The implementation was introduced in PR #24518 by @Laughing-q, and it brings semantic segmentation into the same familiar workflow YOLO users already know.

:puzzle_piece: New YOLO26 semantic segmentation models

We’ve added new pretrained YOLO26 semantic segmentation checkpoints with the -sem suffix:

  • yolo26n-sem.pt
  • yolo26s-sem.pt
  • yolo26m-sem.pt
  • yolo26l-sem.pt
  • yolo26x-sem.pt

These models are designed for tasks where every pixel is assigned a class, making them a great fit for applications like:

  • autonomous driving
  • medical image parsing
  • land-cover mapping
  • industrial scene understanding
  • general semantic labeling

You can explore the latest model family on the Ultralytics YOLO26 page.

:building_construction: Dedicated semantic segmentation pipeline

Semantic segmentation is not an add-on here, it’s a native pipeline with dedicated components including:

  • SemanticSegmentationModel
  • SemanticSegmentationTrainer
  • SemanticSegmentationValidator
  • SemanticSegmentationPredictor

That means a cleaner API, task-specific training flow, and a more consistent developer experience.

:framed_picture: Semantic masks in inference results

A new SemanticMask result type has been added, making semantic outputs easier to inspect, visualize, and use in downstream applications.

:books: Dataset and training improvements

:card_index_dividers: Built-in semantic segmentation datasets

This release adds built-in dataset configs and docs for:

  • Cityscapes
  • Cityscapes8
  • ADE20K

It also adds support for:

  • PNG mask datasets
  • polygon-based semantic training from existing segmentation labels

This lowers the barrier to getting started with real-world semantic segmentation datasets.

:repeat_button: Augmentations now support semantic masks

Training augmentations were upgraded to properly handle semantic masks, including:

  • Mosaic
  • MixUp
  • CutMix
  • flipping
  • perspective transforms
  • letterboxing

This is a big quality-of-life and training-quality improvement for segmentation workflows.

:outbox_tray: Export, benchmark, and deployment updates

Semantic segmentation models now work much more smoothly across export and deployment flows.

Highlights include:

  • semantic task support in export paths
  • benchmark coverage for semantic models
  • backend handling improvements for semantic outputs
  • TensorFlow-specific semantic output support
  • automatic disabling of nms=True for semantic models, since NMS does not apply here

This makes semantic models more practical not just for research, but also for real deployment workflows.

:hammer_and_wrench: Improvements and Fixes

:straight_ruler: Correct FLOPs reporting for FP16 and BF16 models

Half-precision models could incorrectly report 0.0 GFLOPs due to a dtype mismatch. That’s now fixed in PR #24504 by @ahmet-f-gumustas.

:chart_increasing: Final evaluation logging fix

Best metrics were overwriting last-epoch metrics during final_eval in some logging setups. This has been resolved in PR #24530 by @deependujha.

:eyes: Cleaner validation visualizations

Validation plots now respect show_labels and show_conf, making crowded scenes easier to inspect. This improvement landed in PR #24305 by @K-saif.

:package: Better dependency compatibility

Protobuf-related dependency handling was improved to reduce compatibility issues with TensorFlow and Ray Tune in PR #24509 by @onuralpszr.

:toolbox: DALI installation docs for CUDA 13.x

Installation guidance for NVIDIA DALI with CUDA 13.x was added in PR #24498 by @onuralpszr.

:gear: OpenVINO benchmark updates

OpenVINO YOLO26 benchmarks were updated with Intel Core Ultra X7 358H results in PR #24512 by @lakshanthad.

:memo: Documentation updates

A number of docs improvements also shipped in this release:

There was also a dependency bump for idna in the RT-DETR ONNX Runtime example via PR #24540 by @dependabot[bot].

:raising_hands: New Contributor

A warm welcome to @deependujha, who made their first contribution with PR #24530 :tada:

:play_button: Try it now

You can upgrade with:

pip install -U ultralytics

And start exploring semantic segmentation with the same familiar API:

from ultralytics import YOLO

model = YOLO("yolo26n-sem.pt")
results = model.predict("image.jpg")

For a streamlined training and deployment workflow, you can also explore the Ultralytics Platform, which makes it easy to annotate datasets, train models, export, and monitor results.

:link: Release links

Dive into the full details in the v8.4.52 release page, or browse everything included in the full changelog from v8.4.51 to v8.4.52.

:speech_balloon: Feedback

We’d love for you to try v8.4.52, especially the new YOLO26 semantic segmentation workflow, and share your feedback with us. Your testing, bug reports, and ideas help make YOLO better for everyone :raising_hands: