Platform capabilities reviewed on September 19, 2026

September 12–19: assistants, 3D and models

• Choose OpenAI API, ChatGPT via Codex, Claude API or Claude Code in Settings → AI Assistant. The annotation button, connection panel and image-batch workflow follow the selected assistant. API connections work on web and desktop; CLI connections require the desktop app and the corresponding installed executable.
• Experimental 3D point-cloud semantic segmentation: import PLY/PCD, select points with lasso or brush, assign classes, undo/redo, save and export. Desktop/web needs WebGL2. No cuboids, sequences, mobile 3D, 3D AI, quality, splits, dataset versions or evaluation.
• ECDet-L and ECDet-X join the curated detection catalog. Both ECDet and ECSeg now offer S/M/L/X. The auto-annotation reference is generated from the released application catalog. ECDet-L/X checks used a small native/browser sample, not full COCO AP or mobile validation.
• English, Spanish and Brazilian Portuguese cover the application interface, including model information, annotation controls and import/export feedback. Choose the language in Appearance settings; the marketing documentation is currently in English.
• Expanded samples and tutorials, including a labelled 3D Scans sample; reviewed batch drafts and resumable image batches remain separate from manual annotation and local prediction engines.

This is a source-capability review, not a store release announcement. Deployed web code and installed store builds may have different feature availability. Check the version you are using before relying on a new workflow.

Collection and video

• Reusable capture sources: webcam, supported screen sharing, playable local video files and, on Windows and macOS, RTSP/HTTP network cameras decoded by the bundled FFmpeg; source tests, connection status and recovery messages.
• Staged photos and silent clips; manual, timelapse, motion and Smart collection; provenance and source/reason filters.
• Video prediction precomputation and explicit acceptance alongside existing timeline annotation and frame extraction. Prediction sampling is not real-time tracking.
• HEIC/HEIF, TIFF and GIF still-image decoding, platform codec checks and presentation-timeline validation. Animated or multi-page input uses the supported first still; edited output and untouched originals have different export semantics.

Models, annotation and measurement

• Local model evaluation on reviewed test images with task-specific metrics and saved scope identities.
• Experimental ML connections and Pipelines with a versioned TAO RT-DETR runner recipe, job history, logs, artifacts and inference previews.
• RF-DETR Seg model variants and compatible custom segmenter parsing; guided model checks, label mapping and local test inference.
• Text-prompt engine naming, prompt configuration and calibration improvements; batch progress and results, pending-review navigation, cancellation and model-switch safeguards.

Dataset and project operations

• Portable label JSON with prompts, attributes and hierarchy; map labels on import. Keypoint schemas use skeleton templates instead.
• Project duplication and conversion previews, clearer project creation and import workflows, combined dataset exports and split-preserving archives.
• Image redaction, editor controls, media preparation and safer archive/timing validation; YOLO format eligibility and export fixes.
• Mobile capability limits and readiness checks; AI Assistant connection diagnostics; sample projects, tutorial library, download guidance and refreshed product pages.

Capabilities still outside the working loop

Watched-folder adapters are not implemented. Geti Tune is an integration proposal, not a shipped training backend. ML GPU execution needs a separately verified environment; runner predictions are not connected to Models or AI Review. Source changes, deployed web code and store releases have separate delivery timelines.