New Release: Ultralytics v8.3.180

Discussion
, ,
1

Title: Ultralytics v8.3.180 – Faster SAM/SAM2 pipelines, unified tensor outputs, smarter prompts, and better downloads

Summary
Ultralytics v8.3.180 brings a streamlined, feature-based inference path to SAM and SAM2 with standardized torch.Tensor outputs, improved prompt scaling, and a cleaner API. Plus, better dataset downloads, clearer docs, and strengthened CI. Faster to integrate, easier to use, and more consistent end-to-end.

New Features

  • SAM/SAM2 feature-based inference
    • Added a new inference_features API to run decoding directly from image features, avoiding full image re-encodes for each prompt. Great for caching encoder outputs and reducing latency and compute.
    • Public inference_features returns masks and xyxy boxes with scores for convenient downstream use.
    • Shared feature-based logic across SAM and SAM2 removes duplication and simplifies maintenance.
  • Unified outputs and improved prompts
    • Masks and scores now return torch.Tensor for consistent PyTorch-first workflows.
    • Prompt handling improved: _prepare_prompts now accepts src_shape and dst_shape to precisely scale points/boxes across pipelines and image sizes.

Improvements

  • Documentation and tutorials
    • Corrected SAM Predictor example label format to prevent common mistakes.
    • Added a Comet ML tutorial video showing YOLO11 integration to speed up experiment tracking and iteration.
  • Downloader UX and reliability
    • Multi-threaded downloads now show progress and clearer logs.
    • Removed curl dependency for VOC/VisDrone datasets, improving cross-platform robustness.
  • CI and infrastructure
    • Updated actions/checkout to v5 for improved reliability.
    • SBOM generation hardened with a dedicated virtual environment.

Bug Fixes

  • Fixed Anchore SBOM GitHub Action reliability.
  • Fixed minor formatting in CI.md.

What’s Changed

  • Fix Anchore SBOM action by @glenn-jocher in Fix Anchore SBOM action
  • Fix CI.md table by @glenn-jocher in Fix CI.md table
  • Add YouTube tutorial to docs by @RizwanMunawar in Add https://youtu.be/hfSK3Mk5P0I to docs
  • Fix SAM Predictor example in docs by @Laughing-q in Fix SAM Predictor example in docs
  • Progress multi-thread downloads by @glenn-jocher in Progress multi-thread downloads
  • Bump actions/checkout from 4 to 5 by @dependabot[bot] in Bump actions/checkout from 4 to 5 in /.github/workflows
  • New inference_features for SAM models by @Laughing-q in ultralytics 8.3.180 new inference_features method for SAM models

Why it matters

  • Faster pipelines: Decode multiple prompts from cached features for lower latency and compute.
  • More robust code: Standardized tensor outputs simplify batching, devices, and downstream processing.
  • Better accuracy: Explicit prompt scaling improves segmentation quality across varied image sizes.
  • Cleaner APIs: Less duplication and more consistent methods make custom workflows easier to build and maintain.
  • Smoother UX: Clearer docs, progress-aware downloads, and dependable CI save time and reduce friction.

Getting Started

  • Release notes: Ultralytics v8.3.180 Release
  • Full changelog: Compare v8.3.179…v8.3.180

We’d love your feedback
Please upgrade to v8.3.180, try the new SAM/SAM2 inference_features workflow, and let us know how it impacts your pipelines. Share benchmarks, edge cases, or ideas in Discussions—your input helps the community and the Ultralytics team keep improving.

Huge thanks to the contributors who made this release possible, and to the YOLO community for the ongoing support and insights.