New Release: Ultralytics v8.3.178

Discussion
, ,
1

Title: Ultralytics v8.3.178 – Faster, lighter Docker images, export-ready workflows, and smoother preprocessing

Summary
Ultralytics v8.3.178 delivers a major Docker overhaul with a clear split between lightweight inference and full export images, improved CI, headless dependencies for leaner builds, and small API enhancements (LetterBox) for more flexible preprocessing. Expect faster pulls, cleaner installs, and more reliable export paths.

New Features

  • Docker image split for clarity and speed
    • New latest-python image for slim inference workflows.
    • New latest-python-export image with full export toolchain (Edge TPU, NCNN, IMX, Paddle) for production-ready conversions.
  • Preprocessing flexibility
    • LetterBox now supports padding_value and interpolation for finer control over resize/pad behavior across diverse data types.

Improvements

  • CI and image taxonomy
    • CI auto-builds derivative images (runner, jupyter, cpu, python-export) from base images; unified push flow.
    • Tests and benchmarks validate export paths against latest-python-export.
    • Image descriptions and tags clarified; consistent YOLO naming and handling across GPU and Jetson variants.
  • Smaller, more compatible images
    • Switched to opencv-python-headless and default-jre-headless in Docker builds.
    • Consolidated RUN steps, added cleanup and security updates for fewer layers and faster builds.
  • Export and backend reliability
    • NCNN installs now use --no-deps to avoid pulling opencv-python; mirrors added in autobackend loader.
    • Docs/comments updated with RKNN export artifact name yolo11n_rknn_model.
  • Supply chain transparency
    • SBOM generation isolated to a dedicated sbom-env via uv sync; Anchore scans produce SPDX JSON outputs.

Bug Fixes

  • Addressed dependency conflicts in NCNN installs by enforcing --no-deps, reducing unintended pulls and compatibility issues on minimal and ARM64 environments.
  • Resolved GUI-related conflicts in containers by adopting headless packages.

Why it matters

  • Faster default experience: Pull the slim inference image for quick starts; use the export image only when you need conversions.
  • Clearer Docker taxonomy: Easier to choose the right image, with consistent behavior across GPU and Jetson.
  • Leaner, safer builds: Smaller images, fewer layers, and improved security posture.
  • Better preprocessing: Control padding and interpolation to improve model robustness on varied data.
  • Stronger compliance: Cleaner SBOMs help organizations with audits and governance.

What’s Changed (PRs and authors)

  • Scope publish.yml SBOM by Glenn Jocher (@glenn-jocher) – PR: Scope publish.yml SBOM
  • Install NCNN with ‘–no-deps’ by Glenn Jocher (@glenn-jocher) – PR: Install NCNN with ‘–no-deps’
  • Use opencv-python-headless in Dockerfiles by Glenn Jocher (@glenn-jocher) – PR: Use opencv-python-headless in Dockerfiles
  • Support padding_value and interpolation in LetterBox for better compatibility by @Laughing-q – PR: Support padding_value and interpolation in LetterBox for better compatibility
  • Combine Dockerfile RUN commands by Glenn Jocher (@glenn-jocher) – PR: Combine Dockerfile RUN commands
  • Refactor derivative Dockerfile images by Glenn Jocher (@glenn-jocher) – PR: Refactor derivative Dockerfile images
  • ultralytics 8.3.178 new lighter Dockerfile-python images by Glenn Jocher (@glenn-jocher) – PR: ultralytics 8.3.178 new lighter Dockerfile-python images

Links

  • Release notes: Ultralytics v8.3.178
  • Full changelog: Compare v8.3.177…v8.3.178

How to get started

We’d love your feedback
Give the new images a try, especially if you export to NCNN, Edge TPU, IMX, or Paddle backends, or if you’re running on minimal or ARM64 environments. Share your results and suggestions in Discussions, and report any issues on GitHub. Your feedback helps the YOLO community and the Ultralytics team keep improving.