Comparison
Short answer
AnnotateIt is best suited to individuals collecting, finding, annotating and evaluating visual data on their own device: camera capture, local AI, semantic search, reproducible splits and dataset versions in one app. Optional ChatGPT or Claude adds conversational annotation. Choose CVAT for shared annotation jobs, reviewer workflows, 3D data, automatic video tracking and its open-source ecosystem. CVAT Online needs an account, not a Docker installation; self-hosted CVAT can run on your own workstation or server.
Both tools support AI-assisted image and video annotation. The useful comparison is how you collect data, connect models, review results and reproduce a dataset. AnnotateIt stores projects locally and runs its local models on-device; optional OpenAI, Anthropic and ML-runner connections send selected data outside that local workflow.
Annotate with ChatGPT or Claude
Ask for boxes, segmentation polygons or image classes in plain language. Work with one image or video frame in chat, or give Auto-annotate one instruction for a batch of images. Review the results directly in AnnotateIt.
Open Annotate with ChatGPT or Claude above Next. The current image or frame is attached by default; you can turn it off.
Create, rename, recolor or delete project labels with confirmation. Ask follow-up questions to refine pending boxes or polygons.
Review the usual AI suggestions on the canvas. Use the existing Accept / Reject controls; the chat reports when nothing was annotated.
Optional online feature: an attached chat message sends one image or one frame and context to your selected provider, OpenAI or Anthropic. Starting an assistant batch sends eligible images and project labels from your chosen scope. No whole-video or keypoint generation. Local AI tools remain separate.
Four workflows to try with your own data. The detailed comparison below includes CVAT capabilities and external alternatives.
Connect a webcam, use a recording, or open an RTSP/HTTP IP camera on desktop. Select frames manually, by interval, motion or available Smart rules; inspect captures before adding them to the dataset.
Capture sources and limits →Search image content with a text query on your device. Inspect ranked matches and a match heatmap, then use the results as the scope for batch annotation.
Semantic search workflow →Download a curated local model or import a supported ONNX file with a guided test. Alternatively, ask ChatGPT or Claude for annotations and refine its pending suggestions in conversation; this optional connection uses the selected provider, OpenAI or Anthropic.
Conversational annotation →Check dataset integrity, create a seeded split and freeze a version. Run an available prediction engine on reviewed test images and retain the benchmark scope with its metrics.
Local model evaluation →
| CVAT | AnnotateIt | |
|---|---|---|
| Conversational annotation & labels | CVAT has AI Agents and SAM 3 label-based text prompts. The reviewed official sources did not establish the same built-in ChatGPT/Codex conversation for canvas annotations and project labels. Custom integrations may provide related behavior. | Chat and batch image annotation → editable boxes, polygons or image classes → standard Accept / Reject. Auto-annotate supports a shared instruction, up to 2 concurrent API image requests, progress, cancellation and resume in Pending AI Review. Desktop Codex and Claude Code run sequentially. Chat also manages project labels with confirmation. |
| Camera, screen and IP-camera capture | Documented inputs include files, remote URLs, shares and cloud storage. A built-in webcam, screen or live RTSP capture workflow was not found in the reviewed documentation; external capture can produce files for import. | Webcam, supported screen/window/tab sharing, and playable video files. Desktop adds direct RTSP/RTSPS, MJPEG, HLS and HTTP(S) JPEG snapshots through a bundled decoder. IP streams are unavailable in browser/mobile; native macOS and mobile omit screen capture. |
| Smart frame selection and recording | Video import can select a start, stop and frame step. That samples existing media; the reviewed documentation does not establish live motion-, uncertainty- or diversity-triggered collection. | Manual, Timelapse, Motion and Smart capture. Available models supply uncertainty or visual-diversity signals. Camera, screen and network sources can record silent clips when supported. Captures stay staged until Accept. Uncertainty and novelty are review signals, not guarantees of useful training examples. |
| Search by image content | Annotation editor filters are built in. Text-similarity search is available in FiftyOne, which has a documented CVAT integration; it is a separate application and workflow, not the CVAT filter panel. CVAT annotation filtersCVAT integration with FiftyOneFiftyOne text similarity search | On-device semantic image search, ranked results and a coarse match heatmap. Use search results as a batch-annotation scope. Requires a supported encoder and platform; unavailable in native mobile builds. |
| Where it runs and setup | CVAT Online is hosted: create an account. Community and Enterprise are self-hosted options; the server can run on a workstation or separate infrastructure. Docker deployment is not required to use CVAT Online. | Open the browser app or install Windows, macOS, iPhone or iPad. Projects are stored on the device. Local annotation needs no AnnotateIt account or annotation-server deployment. |
| Offline use and data flow | A suitably configured self-hosted installation can keep data and models in your environment, including on the same computer. CVAT Online uses hosted infrastructure; selected external models and storage have their own data flows. CVAT self-hosted installation guideCVAT model integration by edition | Local annotation and local model inference stay on-device. Cache the web app/models first; native builds bundle default engines. Optional ChatGPT or Claude sends previews/context to the selected provider (OpenAI or Anthropic), and optional ML jobs send selected datasets to your runner. |
| Native apps and platform limits | The annotation interface is a browser client connected to a CVAT server or CVAT Online. | Desktop adds direct IP-camera decoding and a loopback API. Mobile retains the local annotation/dataset workflow with smaller built-in models; heavy downloads, custom ONNX, semantic search, external AI Assistant and 3D editing are unavailable there. |
| Accounts and access control | Accounts, organizations and role-based team access. Self-hosted authentication is part of the deployment. | No AnnotateIt account for local annotation. External AI API/CLI access belongs to your selected provider account. Local project organization does not provide a shared multi-user annotation service. |
| Assignments and human review | Tasks/jobs, assignees and reviewer workflows; quality-control capabilities vary by edition. Suitable for a shared annotation operation. | Review local AI drafts and manually correct annotations. Single-user work: no shared queue, simultaneous editing or reviewer assignment service. |
| Ground truth, honeypots and consensus | Ground-truth comparison, hidden validation frames, feedback and consensus replica/merge workflows are documented. Availability depends on the edition and plan. | Pending AI Review and dataset-integrity checks. No equivalent multi-annotator consensus, honeypot assignment or formal adjudication workflow. |
| Quality requirements and completion feedback | Online/Enterprise document hierarchical rules with per-subset thresholds and attribute comparisons. Version 2.74 adds requirement configuration and reports; 2.75 adds Jaccard/Dice and micro/macro/worst-label aggregation. Immediate feedback uses configured validation data CVAT quality requirementsCVAT 2.74.0 — August 26, 2026CVAT 2.75.0 — September 10, 2026CVAT ground truth and honeypots | Integrity rules and separate model-evaluation metrics. No annotator-versus-ground-truth requirement tree or quality gate on assigned-job completion |
| Audio intervals | The 2.76 release and tagged source include an audio interval workspace with snapping and waveform navigation. Confirm availability in your deployment; the general product overview still concentrates on visual annotation | No audio annotation project or waveform-interval editor; recorded capture clips are silent |
| Layers and mask editing | Layer visibility and z-order controls; 2.74 changes to per-layer visibility with Shift-assisted toggling. Mask editing includes local undo/redo, with further fixes in 2.75 CVAT annotation layersCVAT 2.74.0 — August 26, 2026CVAT 2.75.0 — September 10, 2026 | Annotation ordering, visibility and undo/redo exist. These do not establish parity with CVAT layer-group visibility or every mask editing operation |
| Supported tasks and 3D | Image/video detection, segmentation, keypoints and frame tags; 3D point-cloud annotation with cuboids. Flexible shape types within tasks. | 2D detection, instance segmentation, skeleton keypoints and single-label, multi-label or hierarchical classification. Also includes experimental PLY/PCD point-cloud semantic segmentation on desktop/web with WebGL2; no cuboids, sequences or sensor fusion. |
| Shapes and attributes | Boxes, rotated boxes, polygons, polylines, points, ellipses, masks, cuboids, skeletons and tags; object attributes. Version 2.75 adds point-based rotated rectangle/ellipse drawing. Both products have substantial manual editing tools. CVAT editions and toolsCVAT dataset formatsCVAT 2.75.0 — September 10, 2026 | Boxes, circles, polygons, open polylines, brush masks and skeletons, with text/number/select/checkbox/radio attributes. Imported rotated boxes remain editable. Preservation depends on the export format. |
| Interactive AI and text prompts | Detectors, interactors and trackers. SAM 3 supports label-name text prompts combined with a visual example to find matching objects; this is not limited to one-click segmentation. Availability depends on edition and plan. | Local MobileSAM, SAM 2.1, SAM 3 Tracker for single-image masks, Grounding DINO, CLIP/SigLIP 2 and RTMPose where supported. Text-prompt and visual-example tools operate locally; optional ChatGPT or Claude is a separate online mode. |
| Connecting custom models | Nuclio functions, SDK/CLI auto-annotation, and native AI Agents; Online/Enterprise also support documented third-party integrations. An agent is a model process that answers annotation requests, not necessarily a chat agent. CVAT model integration by editionCVAT auto-annotation SDKCVAT AI Agents | An experimental ONNX import wizard inspects a supported graph, maps classes and requires a successful test. Inference runs on-device. Supported architectures include YOLOv8 tasks, RT-DETR/DETR, compatible segmenters and classifiers; arbitrary ONNX graphs are not guaranteed. |
| Ready-to-use models | Ready models and model integrations are available. The model catalog and setup requirements vary between hosted and self-hosted editions. | Curated detector/segmenter downloads include D-FINE, RT-DETR variants, DEIM, RF-DETR and EdgeCrafter families. Map labels and test before use. The model catalog and runtime requirements are documented separately. |
| Batch auto-annotation | Automatic annotation processes task data with a selected model. Supports label mapping and configurable options such as thresholds; detector ROI can restrict inference to part of a frame. | Run supported local engines over a dataset, selection, filter or semantic-search results. ChatGPT batch accepts one instruction for eligible still images, with progress, cancellation/resume and pending review. ChatGPT requests use your external account and limits. |
| Video interpolation and model tracking | Keyframe interpolation and automatic tracking. SAM 2 tracking is documented through Online/Enterprise AI Agents and Enterprise Nuclio deployment. It follows existing masks/polygons across frames. | Keyframe box/object-mask tracks, compatible-polygon interpolation, frame keypoints and local frame extraction. Model predictions can be sampled/cached and accepted, but no model follows an object identity automatically. The SAM 3 Tracker model name does not imply video tracking in AnnotateIt. |
| Video classification ranges | Frame tags can be propagated over a range; the linked source preserves tag object types during range copying. | A dedicated range-labeling workflow: select start/end and assign a class across that span. It is stored as per-frame classification for export and versioning. The difference is the interaction, not exclusive support for multi-frame labels. |
| Annotation filters and keypoints | Filters combine label, shape/type, geometry and attributes. Version 2.75 adds source filtering (auto, semi-auto, manual, file, consensus); multiple skeleton instances per frame are supported. CVAT annotation filtersCVAT 2.75.0 — September 10, 2026CVAT editions and tools | Filters include geometry size, prediction score and annotation source with and/or/nor logic over current edits. Multiple skeletons per image; local RTMPose can propose 17 COCO joints from a person box. |
| Dataset integrity without a reference annotation set | CVAT QA compares work with validation data. External Datumaro can also validate missing labels, annotation anomalies and distributions, so integrity checks are available in a CVAT-based pipeline. | Built-in local scan for unreadable/missing media, invalid or out-of-bounds geometry, required attributes, exact duplicates, unusual dimensions, class imbalance and video coverage. No separate GT job is needed for these checks. |
| Train/validation/test splits | Subsets are assigned to tasks within a project. External Datumaro provides seeded and task-aware split transforms; CVAT users can prepare splits outside the editor. | An in-app deterministic image split planner with seed, label stratification and locked assignments. Exact duplicate content stays in one subset. This does not prevent every form of temporal, subject or near-duplicate leakage. |
| Immutable versions and activity history | Task/project exports and backups, annotation editing history and event/analytics tooling. External tools can extend dataset version management. CVAT task and project backupCVAT quality controlDatumaro dataset operations | Freeze dataset media, annotations, labels, tracks and split state as a version; compare, export or restore it. A separate activity history records local dataset operations. It is not an identity-based multi-user audit log. |
| Model evaluation | Annotation QA and model evaluation are different tasks. CVAT exposes data through SDK/integrations, including FiftyOne, for external model-analysis pipelines; the reviewed QA dashboard is not a local ONNX benchmark runner. | Run an eligible prediction engine on reviewed test images and retain the benchmark scope with task-specific/per-label metrics. Native video timelines and unsupported text/visual/ChatGPT engines are excluded. |
| Training and deployment | Connect annotation to an external training stack through APIs, exports and integrations. Model-serving functions for annotation do not by themselves mean CVAT trains or deploys your production model. | Experimental desktop Pipelines connect to a separately prepared TAO RT-DETR runner for supported image-detection fine-tuning and prediction preview. Requires its own infrastructure; not general hosted training or a production endpoint. |
| Multiple datasets and exports | Projects group tasks under a shared label schema. Standard exports and external Datumaro operations support dataset transformations and merging. | Independent datasets with separate splits, versions and history. Combine selected image datasets into an export while preserving, recomputing or removing splits. |
| Formats and round trips | Broad import/export support, including CVAT XML, Datumaro, COCO and YOLO; supported shapes and metadata vary by format. | COCO, YOLO, Pascal VOC, Datumaro, MOT/MOTS, KITTI, Supervisely Video and ZIP where applicable. An AnnotateIt sidecar preserves extra metadata for re-import here; other tools may ignore it. CVAT XML is not an AnnotateIt import format. |
| Image preparation and redaction | Image rotation and display adjustments assist annotation. A comparable built-in editor for saving cropped/resized/redacted image pixels was not established by the reviewed docs; external preparation remains possible. | Crop, resize, rotate, flip, color adjustment, blur/pixelation/solid-fill redaction within the dataset workflow. Pixel edits require checking affected annotations. Redaction tools alone are not a compliance guarantee. |
| Portable project backups | Download a task/project ZIP containing data, settings and annotations and recreate it from backup; portability does not depend on a server database dump. | Project archives plus whole-profile backups containing projects and downloaded models, with content-hash media deduplication. Restore between supported installs; maintain a separate backup beyond local dataset versions. |
| Automation and ecosystem | Server REST API, SDK, CLI and webhooks. Version 2.76 adds server-level user/organization webhook events; 2.75 removes older update-payload fields. Version 2.74 adds S3-compatible backing storage for video tasks. Deployment and permissions apply. CVAT auto-annotation SDKCVAT integration with FiftyOneCVAT 2.74.0 — August 26, 2026CVAT 2.75.0 — September 10, 2026CVAT 2.76.0 — September 16, 2026 | Desktop loopback REST API with access tokens for scripts on the same machine. It is not a shared network annotation service. |
| Cost and source availability | Community is free and open source. Online has free/paid plans; Enterprise is a separate offering. Check current plan entitlements for hosted models and QA. | The application is currently free with unlimited projects; application source is not public. OpenAI/Anthropic API usage, CLI account access and optional runner infrastructure have separate costs or limits. |
Capabilities and working conditions
Current CVAT capabilities, edition boundaries and the practical situations behind the comparison.
SAM 3 label-based prompts combine a class name with a visual example to segment other instances. AI Agents connect custom model processes to annotation requests. SAM 2 already provides automatic video tracking in supported deployments.
The August–September releases add hierarchical quality requirements, source filters, rotated drawing, layer visibility changes, S3-backed video tasks and server webhooks. Audio interval tools also appear in the tagged release. Ground-truth QA, consensus and FiftyOne integration are longer-standing capabilities, not features first released on this review date.
CVAT SAM 3 text promptsCVAT AI AgentsCVAT SAM 2 trackingCVAT ground truth and honeypotsCVAT consensusCVAT integration with FiftyOneCVAT 2.74.0 — August 26, 2026CVAT 2.75.0 — September 10, 2026CVAT 2.76.0 — September 16, 2026
Community provides a self-hosted open-source annotation system. Online provides the hosted service. Enterprise is the commercial self-hosted offering. Free Community is not the same product configuration as an Online free account.
The model-integration documentation lists Nuclio for Community/Enterprise and selected hosted functions for Online. Native AI Agents and third-party model connections are documented for Online/Enterprise. The SAM 3 release describes a hosted evaluation mode and production availability in paid Online/Enterprise; use the linked current entitlements when choosing.
CVAT editions and toolsCVAT model integration by editionCVAT SAM 3 text prompts
AnnotateIt connects supported sources, selects candidate frames and keeps them in a staging gallery before import. Source/session/capture-reason metadata helps revisit how examples were collected. Watched-folder ingestion is not implemented.
For existing datasets, local semantic search can lead straight to batch annotation. A CVAT-based workflow can use an external recorder or FiftyOne for collection and curation. The advantage here is an integrated workflow on one device.
CVAT tasks and data sourcesCVAT integration with FiftyOneFiftyOne text similarity search
CVAT Online can be a straightforward choice for one person too: it does not require maintaining Docker. Self-hosting gives more infrastructure control but also adds administration.
AnnotateIt is useful when you want local collection, model execution, correction, splits, versions and evaluation in a standalone app. If collaborators need assignments and simultaneous access to the same project, CVAT provides a team layer that AnnotateIt does not.
For local annotation, AnnotateIt runs in a browser or native app without an annotation-server installation. Downloadable models still need compatible hardware and an initial download.
CVAT Online also avoids local server installation. Self-hosted CVAT runs a server stack, potentially on the same workstation. AnnotateIt optional training is separate: its experimental ML runner still needs its own prepared environment.
Both a local AnnotateIt workflow and an appropriately configured self-hosted CVAT workflow can keep data under your control. Self-hosted CVAT does not inherently require sending it to a different computer or vendor cloud.
AnnotateIt local tools avoid an annotation server. Optional ChatGPT or Claude/API requests send image previews and context to the selected provider (OpenAI or Anthropic); runner jobs transfer selected data to the configured runner. Choose the local workflow when those transfers are unsuitable. Local storage does not itself satisfy company approval, access-control or certification requirements.
CVAT self-hosted installation guideCVAT model integration by edition
AnnotateIt combines integrity checks, seeded image splits, immutable versions and eligible model benchmarks. Exact duplicates are grouped within a split; related people, sessions and near-identical video frames still require your own leakage checks.
CVAT has task subsets and portable backups. Datumaro and FiftyOne can extend a CVAT pipeline with dataset preparation and analysis. AnnotateIt makes these local steps accessible together; it does not replace CVAT consensus or formal reviewer management.
CVAT projectsCVAT task and project backupDatumaro split transformsCVAT integration with FiftyOne
CVAT offers a hosted or self-hosted annotation system with team workflows, extensive shape support, automatic tracking and model integrations. AnnotateIt offers local collection, semantic search, on-device models, review, dataset preparation and evaluation, plus optional ChatGPT or Claude annotation. Choose based on the workflow and deployment you need; neither a model count nor a feature checklist establishes better annotation accuracy.
Migration guide
Assignments, reviewer decisions, consensus replicas and job history are not transferred as a shared workflow. Use a small representative round trip before migrating a large dataset.
Yes. It documents native AI Agents, other model integrations and SAM 3 label-name text prompts with visual examples. These features provide model-assisted annotation.
The reviewed official material did not establish an equivalent built-in ChatGPT/Codex conversation that edits pending annotations and manages labels. Custom API, SDK or third-party integrations may provide related behavior. AnnotateIt includes that optional online conversation and a separate still-image batch workflow.
The reviewed documentation describes importing local/remote files and cloud data, not a built-in live webcam or RTSP capture interface. External capture can feed a CVAT pipeline. AnnotateIt desktop connects to supported IP-camera streams directly and stages selected frames for review.
CVAT has annotation and metadata filters. FiftyOne supports text-similarity search and integrates with CVAT, so this is available through a separate workflow. AnnotateIt includes on-device semantic search and batch annotation from the result set on supported platforms.
Yes: tags can be propagated across frames. AnnotateIt provides its own start/end range interface. Multi-frame tagging is not unique to AnnotateIt.
No. AnnotateIt is currently free to use, but its application source is not public. CVAT Community is open source. Free use and source availability are separate questions.
Not simultaneously through a shared service. You can transfer project archives, but there is no shared queue, locking, merge or multi-annotator consensus. CVAT is the stronger fit for those requirements.
No. It supports keyframes, interpolation and separate model predictions, but no model follows an object identity through the video. SAM 3 Tracker is used for single-image masks here. CVAT documents automatic SAM 2 tracking for supported Online/Enterprise configurations.
No. Local annotation and local models operate on your device. ChatGPT and Claude are optional online connections that send selected image previews and context to OpenAI and Anthropic respectively. Windows ChatGPT sign-in requires Codex and eligible account access; API-key usage has separate billing.
Its experimental desktop Pipelines integration can send a frozen supported image-detection dataset to a separately prepared TAO RT-DETR runner. You operate that environment. It is not a universal training service or production deployment platform.
Use a mutually supported format such as COCO, YOLO, Pascal VOC or Datumaro, according to the annotation types. Validate a representative export/import first: application-specific history and all metadata do not automatically transfer.
AnnotateIt ships ready-made, fully annotated sample projects. Open one and click around — the annotator, the AI tools, the export flow — before importing anything of your own.
Create a local project without an AnnotateIt account. Projects stay on your device; optional external AI and ML connections have their own data transfers and account requirements.