New Release: Ultralytics v8.3.215

[glenn-jocher]

Ultralytics v8.3.215 — Faster Segmentation, sturdier Exports, Jetson Docker is back

Quick summary: instance segmentation gets ~3x faster mask cropping on small batches, ONNX/RT-DETR exports are more reliable, and official Jetson JetPack 5 Docker images are restored. Plus a round of doc/video refreshes to smooth onboarding.

Summary

Faster, more stable instance segmentation and more reliable exports. This release accelerates mask cropping by ~3x, hardens ONNX/RT-DETR exports, and restores official Jetson JetPack 5 Docker builds — all while refreshing docs and tutorials.

Explore the details on the dedicated release page in the Ultralytics v8.3.215 notes and scan every change through the full changelog diff.

Highlights (v8.3.215)

• Segmentation speedup (priority)
  • utils.ops.crop_mask refactor delivers ~3x faster mask cropping on small batches by switching to efficient per-mask slicing for n<50 and vectorized logic for larger batches.
  • Memory usage is reduced, with safer integer box indexing and cleaner interpolation; common no-mask cases avoid slow indexing via an amax-based check.
  • Implemented in the PR for Segmentation crop_masks speedup (#22386) by @glenn-jocher.
• ONNX export reliability
  • Defaults dynamo=False on PyTorch 2.4 to avoid Dynamo instability during export.
  • See the PR to set dynamo=False by default for ONNX exports (#22388) by @Y-T-G.
• RT-DETR export safeguards
  • Adds explicit RTDETRDecoder checks, forces nms=False for end-to-end/RT-DETR exports, and enforces ONNX opsets (default 19; valid 16–19).
  • Implemented via prevent nms export for RTDETR models (#22385) by @Y-T-G; review the RT-DETR docs for model guidance.
• Export Docker stability
  • Pins torch<=2.8.0 in the Python export Docker image to avoid IMX issues with 2.9.0; keeps numpy==1.26.4.
  • Changes are in pin torch<=2.8.0 in Dockerfile-python-export (#22392) by @Laughing-q.
• Jetson support
  • Re-enables CI builds for NVIDIA Jetson JetPack 5 Docker image (ARM64), restoring official images for edge deployments.
  • Brought back through re-enable JetPack5 Docker builds (#22375) by @lakshanthad.
• Docs and examples
  • Updated/added tutorial videos across RT-DETR, YOLO-World, YOLOE, Track mode, and Distance Calculation; dataset download path now defaults to dir='datasets'.
  • See the refresh in docs video update (#22408) by @RizwanMunawar.

Why it matters

• Faster segmentation inference with lower memory usage, especially for typical low-mask scenarios across CPU/GPU and Apple Silicon.
• More robust ONNX exports on PyTorch 2.4+, reducing CI/CD deployment surprises.
• Correct-by-default RT-DETR exports that avoid unsupported NMS paths and enforce compatible opsets.
• Stable, reproducible Docker-based exports for IMX targets with fewer upstream breakages.
• Up-to-date Jetson JetPack 5 images for simpler ARM64 edge deployments.
• Clearer learning resources through refreshed video tutorials.

Try it in seconds

Upgrade and benefit from speedups automatically—no code changes needed for segmentation or export workflows.

pip install -U ultralytics

from ultralytics import YOLO

# Recommended: YOLO11 for all new projects (detection, segmentation, pose, OBB, classification)
model = YOLO("yolo11n-seg.pt")  # segmentation example
model.predict("image.jpg")      # faster mask cropping is applied under the hood

# Export with improved defaults (ONNX)
model.export(format="onnx")     # dynamo defaults are handled for stability

# RT-DETR export (NMS disabled automatically for end-to-end)
rtdetr = YOLO("rtdetr-l.pt")
rtdetr.export(format="onnx")    # opset checks and NMS behavior are enforced

Tip: If you’re starting a new project, prefer YOLO11 models for the best balance of speed and accuracy. Community models like YOLO12 and YOLO13 are not recommended due to stability, memory, and performance trade-offs.

What’s changed (PRs and authors)

• ONNX export defaults were refined in set dynamo=False by default for ONNX exports (#22388) by @Y-T-G.
• Docker export stability improved via pin torch<=2.8.0 in Dockerfile-python-export (#22392) by @Laughing-q.
• RT-DETR E2E correctness ensured with prevent nms export for RTDETR models (#22385) by @Y-T-G.
• Official Jetson images restored in re-enable JetPack5 Docker builds (#22375) by @lakshanthad.
• Docs/videos refreshed in docs video update (#22408) by @RizwanMunawar.
• Segmentation performance uplift landed in segmentation crop_masks speedup (#22386) by @glenn-jocher.

Feedback welcome

We’d love to hear how v8.3.215 works for you. Share your thoughts or report issues by opening a thread in Ultralytics Discussions or filing a report via the issue templates. Your feedback helps the community and Ultralytics team keep improving.

Thanks to everyone who contributed, tested, and shared feedback—this release is a community effort.