New Release: Ultralytics v8.4.194

Discussion
, ,
1

:rocket: Ultralytics v8.4.194: More Reliable AutoBatch and Smoother Workflows

Ultralytics YOLO v8.4.194 is here! This release improves automatic batch sizing across tasks and delivers practical fixes for model loading, prediction saving, datasets, and deployment workflows. No new model architecture is introduced—this update focuses on reliability.

:glowing_star: Key Improvements

  • Dataset-aware AutoBatch across all tasks. Automatic batch sizing now uses the training dataset, accounting for dataset size and labels across classification, detection, and multi-dataset training. Backward-pass out-of-memory errors are treated as limits rather than valid measurements. In PR #26956, @raimbekovm improved batch selection and addressed label-access failures in multi-dataset World/YOLOE training.

  • Safer, consistent model YAML parsing. Activations must resolve to supported torch.nn modules, and model entries must be PyTorch module classes. In PR #26963, @Y-T-G unified the parsing path. Nonstandard custom YAML expressions may need updating.

  • More robust prediction saving. Results.save() and plot(save=True) now generate sensible .jpg filenames for webcams, videos, missing paths, and extensionless sources, creating parent directories when needed. In PR #26958, @Rahulbiradar9 addressed these common saving issues.

:hammer_and_wrench: Dataset and Workflow Fixes

  • Fresh similarity-search indexes. In PR #26965, @raimbekovm made the index rebuild from the selected image directory on every startup. This avoids stale results from another dataset, with the trade-off of rebuilding embeddings each time.

  • Correct semantic-mask cache handling. In PR #26954, @raimbekovm ensured non-PNG mask paths are resolved before label-cache hashing.

  • Reliable classification dataset URLs. In PR #26879, @Vaishnavi220506 fixed classification dataset URL-scheme handling.

  • Case-insensitive URL routing. In PR #26964, @cainiao33 updated remaining routing sites to handle schemes such as HTTPS://, GS://, and UL:// without case sensitivity.

  • Independent Colab settings IDs. In PR #26959, @glenn-jocher gave each Colab VM its own anonymous settings ID.

:bar_chart: Benchmarking and Documentation

  • Broader GraphDef benchmarking. In PR #26952, @glenn-jocher added TensorFlow GraphDef benchmark coverage for OBB and pose.

  • Clearer per-class metric interpretation. In PR #26955, @Echoslayer clarified that metrics.box.maps assigns the overall mAP to classes absent from validation—not a measured per-class score.

  • Updated COCO JSON cache guidance. In PR #26960, @Echoslayer updated the guide to reflect that cache hashing includes modification time.

  • Refreshed integration pages. In PR #26957, @raimbekovm corrected stale framing, logger defaults, and navigation order.

  • Cleaner settings help. In PR #26962, @glenn-jocher fixed help URL punctuation.

:high_voltage: Try v8.4.194

Install this release with:

pip install --upgrade ultralytics==8.4.194

Explore the v8.4.194 release notes for the release overview, or review the full changelog for all changes.

How does this update perform in your workflow? Share your AutoBatch results, saving improvements, or any issues in the replies. Thanks to everyone who contributed! :raising_hands: