CVAT Digest, October 2025: Smarter Cloud Workflows and Token-Based Automation

Welcome to the October edition of the CVAT Digest, your monthly roundup of the latest features, improvements, and fixes in Computer Vision Annotation Tool (CVAT).

This month’s updates make CVAT smoother for teams working with large cloud-based datasets, 3D projects, and automation workflows. Whether you’re managing storage, creating tasks, or connecting external tools via API, you’ll notice a faster, more flexible, and more secure experience across the board.

New

Token‑based Authentication Across the Platform
You can now use API access tokens everywhere:

• Server API supports API access tokens.
• CLI accepts tokens via the CVAT_ACCESS_TOKEN environment variable.
• SDK can authenticate with an API token via login()/make_client().

Create Tasks Without Labels
Spin up tasks first and finalize taxonomy later. Great for quick intake from cloud or bulk uploads.

Cloud Dataset Handling
CVAT now reliably supports related images for both 2D and 3D tasks from cloud storage, and the Dataset Manifest tool handles 3D datasets across all supported layouts.

Admin Control: Max Jobs per Task
Set a maximum number of jobs allowed per task to keep workloads tidy and predictable.

Configurable Disk‑Usage Health Check
Tune the threshold that triggers the disk‑usage health check to better fit your environment.

Updates

Security & Operations

• Redis upgraded to address a reported CVE.
• FFmpeg 8.0 is now used for modern codec support.
• Helm compatibility restored with Kubernetes pre‑release versions; charts now pull images from the public repo.
• Clarified behavior: CVAT_ALLOW_STATIC_CACHE only affects new tasks (existing tasks keep their configured chunking).
  

SDK & CLI Quality of Life

• The SDK now automatically retries certain transient server errors.
• SDK supports server URLs that include ports (e.g., https://example.com:8443).

Performance Improvements

• Faster task creation from the cloud, even when you don’t have a manifest.
• Lower memory usage when counting objects in tracks during annotation updates and analytics.

Manifest Requirements & Clarity

• Frame width and height are now required in dataset manifests (2D & 3D).
• Manifests can include an optional original_name field, and error messages at task creation are clearer.
• Documentation around supported layouts with related images has been improved.

Cleanup and Deprecations

• You can no longer upgrade directly from releases prior to v2.0.0—plan a staged path through a 2.x release.
• Removed legacy, non‑functional API URL signing code.
• Upcoming change: overly broad filtering of files that merely contain the string related_images is deprecated. CVAT will filter only actual related‑image files according to the input layout.‍

Fixes

• Export integrity: Fixed an issue where tracks could leak between jobs on export.
• Cloud storage workflows: Bulk delete now removes all selected storages; project/task transfers retain the correct storage reference.
• Related images: Detection is more reliable across all supported layouts for both 2D and 3D media.
• UI stability & polish:
  
  • Resolved a crash when loading annotation format metadata.
  • Fixed a sporadic UI error related to reading points.
  • Corrected model card clipping on small displays.
  • Organization description updates now save correctly.
• Server resilience: The backend starts correctly even when analytics are disabled, and a packaging quirk that produced a misleading pkg_resources error message is handled.

Have suggestions or requests for what you'd like to see next? Open an issue on GitHub.