🎉 This release is a big one: YOLOX support is here, bringing a super-fast model under the much more permissive Apache 2.0 license (no more AGPL headaches!). Even more exciting, postprocessing is now fully model-agnostic—so you’re no longer limited to COCO-trained models, but can use your own custom-trained models and class lists with ease. Unlock new use cases and deploy anywhere!

Enhancements

• Added YOLOX support - thanks @aspett (Andrew Pett)!
• Added init/2 callback to the YOLO.Model behaviour for model initialization
• Model-agnostic postprocessing:
  • Removed fixed {8400, 84} shape constraint
  • Now supports dynamic shapes like {batch_size, num_detections, bbox_coords + num_classes}
  • Enables use of custom-trained models with varying class counts
• Optimized NMS implementation:
  • Rewritten using Nx.Defn for 100x performance improvement
  • Added support for variable detection counts and class numbers
• Improved Ultralytics preprocessing:
  • New implementation using :