Geti v3.1.0

🚀 Geti™ v3.1.0 - new models, improved pipeline reliability and much more!

Geti™ 3.1 builds on the same lightweight edge-native architecture of Geti™ 3.0, extending it with several new SOTA models while also improving dataset navigation, inference pipeline monitoring, quantization and other core features.

This release also extends the documentation and it updates some key dependencies (including PyTorch and OpenVINO™) for better performance, stability and security.

To upgrade from an existing Geti™ 3.0 deployment to 3.1, follow the upgrade guide.

🔷 Highlights

• Broader model zoo - enabled the new EdgeCrafter (ECDet) architecture and more RF-DETR variants, giving users more flexibility to choose the best model for their accuracy and speed requirements.
• Filter datasets by subset - the Dataset page and the annotator now offer the possibility to filter media by a specific subset (training, validation or testing).
• Client-provided video files as pipeline sources - upload a video straight from your local filesystem when configuring a source for the inference pipeline, without needing the file to be available in the server beforehand.
• Pipeline observability – it is now easier to monitor the status of the inference pipeline, by means of a health indicator (with fine-grained status info reported for each component) and a one-click 'Test connection' button for sources and sinks.
• Increased UI responsiveness – thanks to HTTP/2, concurrent API requests are handled more efficiently, improving the general responsiveness of the application.

🧠 Models and training

• EdgeCrafter for object detection - the new ECDet (a.k.a. EdgeCrafter) model is now available and fully supported for training, quantization and inference in Geti™. ECDet is available in different size variants: S, M, L, and X.
• New size variants of RF-DETR were added for object detection (N) and instance segmentation (N, 2XL), in addition to the ones already present since 3.0.
• Tiling - All detection and instance segmentation models now support tiling for training and inference, closing the gap for some models that didn't already support it (e.g. D-FINE, RF-DETR, ...)
• Iteration limit for accuracy-aware quantization - a new parameter is available to optionally cap the number of iterations during accuracy-aware quantization, preventing it from running indefinitely.

🗂️ Projects, datasets and annotation

• Project filtering - projects can be searched by name and filtered by task type.
• Pinned active project - any project with a running inference pipeline is now pinned to the top of the list and tagged as "Active" badge for better visibility.
• Annotator media filtering - all media filters are now also available inside the annotator, making it much faster to navigate between images and videos without leaving the annotator.
• Subset filtering - a new filter is available to select media assigned to a specific subset (training, validation, testing or unassigned).

🚀 Deployment and inference pipeline

• Video upload for source creation - it is now possible to create a new source of type 'Video file' by uploading a video from your local filesystem. Previously, the video needed to be already available in the server beforehand, which is not a practical setup for many users.
• Test connection - sources and sinks can be validated before use through the new 'Test connection', with an inline connection-status badge reflecting the result.
• Pipeline health monitor - a new indicator in the 'Pipeline' page aggregates status information about the source, model and sink in real-time, providing informative diagnostic info when something goes wrong in one of these components.

⚙️ Platform, API, and infrastructure

• Upgrade mechanism - in-place migration from a 3.x release to a newer version is now fully supported, automated and documented for all installation modes: Docker, MSIX, Tauri desktop app and scripts (install.sh / install.ps1). All projects and models are preserved, all data will be migrated automatically to the newer version as necessary. As an extra safety mechanism, the database and the application files are backed up before migration, and an informative error message should be displayed in case of problems.
• HTTP/2 REST API - the communication between the UI and the server now leverages HTTP/2, which is more efficient to handle concurrent requests and it enables more flexibility (notably, it allows having more than 7 jobs queued at once). The necessary TLS certificates are generated automatically on installation, regardless of the chosen mode. Note that the URL to access the UI now starts with https:// instead of http://.
• Faster startup - several imports were optimized, noticeably reducing the time to launch the application.
• Pretrained weights provisioning - for some model architectures, it is now possible to automatically download the pre-trained weights from a mirror location if the primary one becomes unavailable, for better reliability. Moreover, for users who need to install Geti in an air-gapped environment, a technical path is available and documented.
• Several dependencies were updated, most notably torch (2.12.1), openvino (2026.3.0) and rfdetr (1.8.0).

🐞 Bug fixes

• Resolved an issue causing training to sometimes fail or hang when tiling is enabled.
• Resolved an issue where trained models would be exported to OpenVINO IR with a static batch dimension of 1, causing problems on batch inference.
• Resolved various issues where Intel® XPU devices would occasionally not be properly recognized by the platform.
• Resolved two issues causing D-FINE models to crash or output poor quality predictions on some Intel® XPU devices.
• Resolved an issue where Mask-RCNN EfficientNet-B2 would crash on inference on some Intel® XPU devices.
• Resolved several issues and race conditions in the inference pipeline that caused instability and silent failures.
• Resolved some issues with the Auto-segmentation annotation tool (SAM), that would often timeout on Tauri / MSIX.
• Resolved various issues with the download of pre-trained weights, which is now more robust thanks to a retry mechanism and mirrors.
• Resolved an issue that prevented the import of some datasets with 16-bit images.
• Resolved an issue that caused errors when importing datasets with out-of-bounds coordinates; the app now sanitizes these values.
• Resolved a visualization issue with intensity mapping configurable parameters, where the displayed decimal precision was insufficient for certain use cases.
• Resolved a formatting issue with training logs.

Download this release

Windows Installer (run Geti as a native Windows application, with prebuilt images for Intel® XPU, NVIDIA® CUDA, and CPU-only environments):
Download CPU-only version installer.
Download Intel® XPU version installer.
Download Nvidia® CUDA version installer

Docker Images

ghcr.io/open-edge-platform/geti-cpu:3.1.0
ghcr.io/open-edge-platform/geti-xpu:3.1.0
ghcr.io/open-edge-platform/geti-cuda:3.1.0

Refer to the installation and upgrade documentation for more details, the guides provide instructions for advanced deployment scenarios too.

---

Geti Library (getitune) v0.2.0

getitune is the low-code transfer-learning library that powers Geti™ training, export, optimization, and inference. This release focuses on widening the model zoo, generalizing the Ultralytics backend to new vision tasks, and making the training and export contracts more predictable.

The getitune library is available as a Pypi package (pip install getitune) or it can be installed from source.

New models

• EdgeCrafter detection model - a new powerful model called EdgeCrafter (a.k.a. ECDet) is now available for object detection, in different sizes (S / M / L / X). This modern model architecture is built on top of a ViT backbone, a RT-DETR-based encoder and a DETR-style decoder. The addition of ECDet to getitune enables users to train this model, quantize it and use it for inference on Intel® XPU.
• More RF-DETR recipes - more size variants of RF-DETR were enabled for object detection (N) and instance segmentation (X, 2XL)

API changes

• Simplified engine creation - engine factory create_engine(...) now accepts either a recipe path or a plain model name (for example create_engine(model="yolox_s", data="path/to/data/root")). It is now unified across all our supported backends.
• Configurable pretrained weights - when instantiating any model (derived from LightningModel), the new parameters pretrained and pretrained_weights allow to optionally supply a checkpoint file to initialize the weights.
• Iteration limit for accuracy-aware quantization: OVEngine.optimize exposes a new parameter max_num_iterations to optionally set a maximum number of iterations for NNCF accuracy-aware quantization.

Other improvements

• Tiling support - tiling is now supported for all detection and instance segmentation architectures.
• Automated benchmarking - all models are now regularly benchmarked through a modern, automated framework that was significantly redesigned and improved. This will help developer prevent regressions and catch issues faster.
• Upgraded PyTorch to 2.12.1, OpenVINO™ to 2026.3 and rfdetr to 1.8.0, resolving several bugs with training and inference with certain devices and model architectures.
• Removed outdated documentation and consolidated the library README around the current API.

Deprecations and breaking changes

• [BREAKING] Discontinued hierarchical classification and rotated detection - these two tasks were removed from the library and from the list of supported task; use flat multi-label classification or instance segmentation as alternatives.

Bug fixes

• The lightning exporter now emits graphs with a dynamic batch dimension instead of a fixed batch size of 1.
• D-FINE models would sometimes crash or output poor quality predictions on Intel® XPU devices.
• Mask-RCNN EfficientNet-B2 would crash on inference on some Intel® XPU devices.

Source code for open source components, provided as required by license.