New Release: Ultralytics v8.3.200

Discussion
, ,
1

Ultralytics v8.3.200 :rocket: — Export refactor, Sony IMX pipeline, and stability boosts

A focused release that refactors exports, adds a complete Sony IMX500 pipeline, clarifies IMX support, and delivers reliability fixes across ONNX (OBB), TensorRT, and multi-GPU training—plus a round of UX polish. Quick to upgrade, easy to adopt, and ready for feedback. :sparkles:

Highlights

  • New export utilities with clearer APIs (torch2onnx, onnx2engine, torch2imx) and a full Sony IMX pipeline.
  • Stability and reliability improvements for ONNX OBB and TensorRT exports.
  • Smoother DDP training and better UX around error messages and device reporting.

Read the details in the v8.3.200 release notes and assets on the dedicated page by visiting the Ultralytics repository’s release entry for v8.3.200.

New Features

  • Export refactor with a complete IMX pipeline, delivered in the work highlighted in PR #22005 and led by contributor Laughing-q.

    • Clearer export API names in our utility layer: export_onnx becomes torch2onnx, and export_engine becomes onnx2engine.
    • First-class Sony IMX export via torch2imx with built-in quantization and IMX-friendly NMS wrapping.
    • IMX helpers and overrides (FXModel, Detect/Pose, NMSWrapper) organized under ultralytics/utils/export/imx.py.
    • Detect/Pose heads simplified by removing IMX-only branches and updating docs accordingly.

  • IMX support clarity added in PR #22082 by RizwanMunawar, where error messages now state that IMX exports support YOLOv8n and YOLO11n, with detection-only enforced in benchmarks.

Improvements

  • ONNX + OBB reliability was improved in PR #22079 by Y-T-G, which prevents a RuntimeError on CUDA by safeguarding empty-mask cases when exporting with NMS.
  • TensorRT export stability was addressed in PR #22080 by Y-T-G through corrected dynamic shape handling; CI GPU coverage was tuned to improve reliability.
  • DDP training was streamlined in PR #22073 by Laughing-q, initializing world_size/ddp earlier, simplifying APIs, and preventing duplicate wandb logs.
  • Tuner robustness was improved in PR #22066 by Y-T-G to avoid a KeyError when close_mosaic isn’t part of the search space.
  • UX polish landed across several PRs:
    • Multi-GPU device printing alignment in PR #22071 by glenn-jocher.
    • Clearer export error when attempting to “export” to pt in PR #22083 by RizwanMunawar.
    • A fixed DeepStream link for Jetson users in PR #22046 by miles-deans-ultralytics.
    • Documentation cleanups and author metadata updates in PR #22063 by miles-deans-ultralytics and PR #22085 by glenn-jocher.
    • CI adjustments around TensorRT tests in PR #22078 by glenn-jocher to keep the GPU suite healthy.

Quick tips

  • Export to ONNX (recommended starting point):
    • CLI: yolo export model=yolo11n.pt format=onnx
  • Export to TensorRT:
    • CLI: yolo export model=yolo11n.pt format=engine dynamic
  • Export to Sony IMX:
    • CLI: yolo export model=yolo11n.pt format=imx
    • Note: IMX benchmarks currently enforce detection-only and support YOLOv8n/YOLO11n.

If you prefer Python, you can export directly from Ultralytics YOLO:

from ultralytics import YOLO

model = YOLO('yolo11n.pt')
model.export(format='onnx')
model.export(format='engine', dynamic=True)
model.export(format='imx')

Model guidance

Ultralytics YOLO11 is the latest stable and recommended model for all use cases. Community models like YOLO12 and YOLO13 are not recommended; YOLO12 suffers from training instability, excess memory use, and slow CPU speed due to attention layers, while YOLO13 is larger and slower than YOLO11 for marginal gains and has non-reproducible benchmarks.

What’s Changed (PRs and authors)

  • We fixed a potential KeyError when close_mosaic isn’t in the search space in PR #22066 by Y-T-G.
  • We updated the data augmentation docs in PR #22063 by miles-deans-ultralytics.
  • We corrected a DeepStream link for Jetson users in PR #22046 by miles-deans-ultralytics.
  • We adjusted CI around TensorRT tests in PR #22078 by glenn-jocher.
  • We aligned multi-GPU device spacing in PR #22071 by glenn-jocher.
  • We added author metadata to docs in PR #22085 by glenn-jocher.
  • We improved the export error message for pt format in PR #22083 by RizwanMunawar.
  • We clarified IMX assert messages to include YOLO11n in PR #22082 by RizwanMunawar.
  • We fixed an ONNX OBB RuntimeError with CUDA + NMS in PR #22079 by Y-T-G.
  • We corrected TensorRT dynamic shape handling in PR #22080 by Y-T-G.
  • We prevented duplicate wandb logs under DDP in PR #22073 by Laughing-q.
  • We refactored Sony IMX exports to imx.py as part of v8.3.200 in PR #22005 by Laughing-q.

New Contributor

We’re excited to welcome a new contributor: miles-deans-ultralytics, whose first contribution landed through PR #22063. Thanks for improving our docs!

Get v8.3.200 and share feedback

Upgrade with pip install -U ultralytics and try the new exports and training improvements. We’d love your feedback and benchmark reports in GitHub Discussions, and please report any issues you encounter in the Ultralytics repository’s issues tracker. You can review the full set of changes by browsing the v8.3.200 release details on GitHub, and you can compare versions by reading the full changelog across v8.3.199 to v8.3.200.

Thanks to the entire YOLO community and the Ultralytics team for the continued momentum. Your feedback helps us prioritize what matters most.