CVAT Vs Label Studio: Which Open Source Labeling Tool Should You Use?

Meta Description: CVAT vs Label Studio compared: licenses, data types, AI-assisted labeling, self-hosting, QA features and pricing, plus which open source tool fits your team.

If your team wants to label training data without paying per seat from day one, two open source tools come up in almost every shortlist: CVAT and Label Studio. Both can be installed on your own servers for free, both have commercial cloud and enterprise editions, and both connect to machine learning models for assisted labeling. They were built for different jobs, though, and that shapes everything from the annotation interface to the export formats.

The CVAT vs Label Studio decision usually comes down to one question: is your data mostly images, video and 3D, or is it a mix of text, audio, documents, time series and images? CVAT grew out of a computer vision team at Intel and is deepest on visual work. Label Studio was designed from the start as a multi-type tool with a configurable interface, and it now leans heavily into LLM evaluation and text tasks.

This comparison covers origins, licenses, data types, AI-assisted labeling, self-hosting, quality control, APIs, export formats, community size and published pricing. If you would rather hand the work to a managed vendor than run a tool yourself, our list of the top data annotation and labeling companies is a better starting point.

What CVAT Is

CVAT (Computer Vision Annotation Tool) is an annotation platform for image, video, audio and 3D point cloud data. According to the company’s own history, Intel engineers Nikita Manovich and Andrey Zhavoronkov started building it in January 2017 as an internal tool, extending the older VATIC video annotation project. The team open sourced it as CVAT 0.1.0 on GitHub in June 2018, under the same GitHub organization as OpenCV, and Intel formally announced it as an open source initiative in December 2019.

In July 2022, CVAT spun out of Intel as an independent company, CVAT.ai, with Manovich and Boris Sekachev as co-founders. CVAT Online, the hosted version, launched in December 2022, and the company also sells managed labeling services. Today the product comes in three forms: CVAT Community (the free, self-hosted open source edition), CVAT Online (cloud subscription) and CVAT Enterprise (self-hosted with commercial features and support).

The open source repository is released under the MIT license.

What Label Studio Is

Label Studio is an open source data labeling tool built by HumanSignal, the company formerly known as Heartex. HumanSignal announced the name change on June 20, 2023, and said both the open source Community edition and the commercial Enterprise edition would stay central to the business. The Label Studio blog says the founders started building it in 2019.

The GitHub project describes Label Studio as a “multi-type data labeling and annotation tool with standardized output format.” It covers images, audio, text, HTML, time series and video, and its labeling interface is defined with XML-style configuration tags, so one install can host very different projects. HumanSignal sells two commercial tiers on top of the Community edition: Starter Cloud for small teams and Label Studio Enterprise, plus a separate data services offering staffed by domain experts.

The Community edition is released under the Apache 2.0 license, with copyright attributed to Heartex.

CVAT vs Label Studio At A Glance

Factor	CVAT	Label Studio
Company	CVAT.ai (spun out of Intel in July 2022)	HumanSignal (formerly Heartex, renamed June 2023)
Open source license	MIT	Apache 2.0
GitHub stars (checked Oct 30, 2026)	About 17,000	About 28,200
Core focus	Computer vision: images, video, 3D point clouds	Multi-type: text, audio, images, video, time series, HTML, PDF
Interface	Fixed, purpose-built annotation editor	Configurable interface built from labeling tags
AI-assisted labeling (free edition)	Serverless models on Nuclio, including SAM	ML backend SDK, with examples for SAM, SAM 2 and Grounding DINO
Review and QA in free edition	Manual review with validation stage and issue tool	No reviewer workflow or agreement metrics
Self-hosting	Docker Compose	pip or Docker
Paid cloud entry price	CVAT Online Solo from $23 per month (yearly)	Starter Cloud from $99 per month
Enterprise pricing	Not published	Custom pricing

Data Types And Annotation Tools Compared

Visual Data And 3D

CVAT is the deeper tool for visual work. Its annotation editor documents rectangles, polygons, polylines, points, ellipses, cuboids, skeletons, a brush tool for masks, and tags. It has a track mode for following objects across video frames and a dedicated 3D task workspace for point cloud annotation.

Label Studio handles images and video too, with rectangles, polygons, ellipses, keypoints, brush masks, a magic wand tool and video rectangles for tracking. Its tag reference does not list a native point cloud or 3D object type, so teams with lidar data generally look elsewhere. If your work is mainly visual, our list of computer vision annotation companies shows which managed vendors work with each type of visual data.

Text, Audio And Mixed Data

This is where Label Studio pulls ahead. Its tag library includes object types for Text, HyperText, Paragraphs, PDF, Table, Chat, Audio, TimeSeries and more, and control tags for named entity labels, relations, taxonomies, free-text answers, ratings, rankers and pairwise comparison. That makes it practical for NER, transcription, conversation review and LLM preference data in one deployment.

CVAT has added audio annotation and lists a Generic TSV format for audio intervals, but text and document labeling are not its focus.

AI-Assisted Labeling Compared

CVAT

In the Community edition, CVAT supports automatic and semi-automatic annotation through pre-built serverless models that run on Nuclio. The repository lists Segment Anything (SAM), Inside-Outside Guidance, RetinaNet, YOLO v7 and others, along with tracker models such as TransT. Turning this on requires an extra Docker Compose component and deploying functions with the nuctl tool, so it is a real infrastructure step rather than a checkbox.

On the paid side, the repository says built-in auto-labeling with SAM 2 and SAM 3 is available in CVAT Online paid plans and CVAT Enterprise. CVAT added AI agents for automated labeling in January 2025.

Label Studio

Label Studio connects to models through its ML backend, an SDK that wraps your model code and runs it as a web server that Label Studio calls. Connected models can pre-annotate tasks for review, return suggestions in real time as annotators click (interactive labeling for rectangles, polygons, keypoints and brushes), and be retrained from annotated data. Example backends include Meta’s SAM, SAM 2 for images, Grounding DINO, MMDetection, OCR engines and spaCy. The documented SAM 2 image backend requires a GPU.

Custom ML backends and predictions from connected models are included in the Community edition. Automated pre-labeling with LLMs, benchmarks and active learning loops are listed for Starter Cloud (as limited or trial features) and Enterprise, and LLM-as-a-judge is Enterprise-only.

Setup And Self-Hosting Compared

Installing CVAT

CVAT’s self-hosted guide uses Docker and Docker Compose. You clone the repository, run docker compose up -d, create a superuser in the server container and open the app on port 8080. The guide documents Ubuntu, Windows 10 through WSL2 and macOS, and Google Chrome is the only supported browser.

Installing Label Studio

Label Studio is lighter to try. A pip install label-studio followed by the label-studio command starts a working server at localhost:8080, and there is an official Docker image and a Docker Compose setup that adds PostgreSQL and Nginx. The docs list PostgreSQL or SQLite as supported databases and recommend at least 8GB of RAM.

Label Studio’s pip install is quicker for a solo test, but a shared deployment of either tool needs a proper database, storage and maintenance.

Collaboration And Quality Control Compared

Free Editions

This is the biggest practical difference between the two free editions. CVAT’s documented manual QA workflow lets an owner or maintainer assign a job to a reviewer in a Validation stage, where the reviewer flags problems with an issue tool, then send the job back to the Annotation stage for correction. That gives small teams a basic review loop at no cost, although review mode does not apply to 3D tasks. CVAT’s docs mark immediate job feedback, quality analytics and the analytics dashboards as CVAT Online and Enterprise features.

Label Studio’s own edition comparison lists no role-based access control, no reviewer assignment and no agreement metrics in the Community edition. The free edition covers the Data Manager, import and export, templates, the tag library, API, SDK, webhooks and ML backends. Teams that need structured review on Label Studio usually move to a paid tier or build review steps around the API.

Paid Editions

CVAT Online Team plans add manual review, ground truth jobs, honeypots, reports and analytics, RBAC, audit logs through the API and batch automatic annotation, with SSO on annual plans. CVAT Enterprise brings them on-premises with dedicated support.

Label Studio Starter Cloud adds RBAC, project-level roles, reviewer assignment, annotation history, limited agreement metrics and a support portal. Enterprise adds customizable permissions, workspaces, custom agreement metrics, annotator dashboards, SSO through SAML or LDAP, SOC 2 and HIPAA compliance in cloud or on-premises, and a 99.9% uptime SLA.

APIs, SDKs And Export Formats Compared

APIs And SDKs

Both tools expose a REST API. CVAT publishes a Python SDK (cvat-sdk) for automating task creation, uploads and exports, and a separate command line tool (cvat-cli). Label Studio publishes label-studio-sdk for creating projects, bulk importing tasks, filtering data and exporting annotations, plus webhooks and the separate ML backend package.

Export Formats

CVAT’s dataset documentation lists more than two dozen formats, including CVAT XML, Datumaro, COCO, Pascal VOC, YOLO, five Ultralytics YOLO variants, segmentation masks, Cityscapes, LabelMe, Open Images, KITTI, MOT, MOTS and ImageNet.

Label Studio’s export guide lists JSON, JSON_MIN, CSV and TSV for any project, COCO, Pascal VOC, YOLO and brush masks as NumPy or PNG for images, CoNLL2003 for text (convertible to spaCy), and ASR manifests for audio transcription. The docs note that Community edition exports from the interface run synchronously, so very large projects can time out, and they point to the Snapshot API, SDK or a console export command instead. Background snapshot exports in the interface are an Enterprise feature.

Community Size Compared

When we checked GitHub on October 30, 2026, Label Studio had about 28,200 stars and 3,700 forks, and CVAT had about 17,000 stars and 4,000 forks. HumanSignal’s homepage says Label Studio is used by more than 1,000,000 AI practitioners. CVAT’s company history says CVAT Online reached 50,000 users in August 2023. Both projects are actively developed and backed by commercial companies. Label Studio’s wider reach partly reflects its broader data type coverage, while CVAT’s base is concentrated among computer vision teams.

Pricing Compared

Both Community editions are free to self-host. You pay for servers, storage, GPUs for assisted labeling and the time to maintain the deployment.

CVAT’s pricing page lists a free CVAT Online plan limited to one project, three tasks and 1 GB of internal storage, with annotation-only export. The Solo plan is listed at $33 per month billed monthly or $23 per month on a yearly plan. Team plans are listed at $33 per user per month billed monthly or $23 per user per month yearly, for 2 to 50 members, with project, task, storage and AI call limits that rise with team size. CVAT Enterprise pricing is not published on the main pricing page.

HumanSignal’s pricing page lists Starter Cloud at $99 per month, with $49 per month for each additional user, for up to 12 users. Label Studio Enterprise and HumanSignal’s data services are custom priced through sales.

At published rates, CVAT Online’s per-user pricing is lower for small hosted teams, but the plans bundle different QA, AI and security features, so compare against the features you need.

Which One Fits Your Team?

Choose CVAT When

Your data is mostly images, video or lidar point clouds and you need tools built for that work: cuboids, skeletons, polylines, video tracks and 3D annotation. It also fits when you need a free review loop, broad vision export formats, or a managed labeling workforce from the same vendor.

Choose Label Studio When

Your projects span text, audio, documents, time series, chat transcripts and images, or you are building preference and evaluation data for LLMs. One deployment can serve NLP, speech and vision teams, and developers get a fast pip install and a Python ML backend for their own models.

When To Use Both Or Neither

Some organizations run CVAT for their vision team and Label Studio for text and audio, since both are free to self-host and export to common formats. Neither is the right choice if you lack the engineering time to host, secure and update a labeling platform, or if you need a large trained workforce more than software. In that case, compare hosted products in our guide to the top AI data annotation platforms or outsource the labeling entirely.

Frequently Asked Questions

Is CVAT Or Label Studio Free For Commercial Use?

Yes. CVAT’s open source repository uses the MIT license and Label Studio’s Community edition uses Apache 2.0. Both permissive licenses allow commercial use, modification and self-hosting. Check the separate licenses on any models you deploy alongside them.

Which Tool Is Better For Video Annotation?

CVAT has the more complete video toolset, with track mode, tracker models and tracking export formats such as MOT. Label Studio’s video rectangle and timeline tags suit simpler video tasks.

Can Label Studio Annotate 3D Point Clouds?

Label Studio’s tag reference does not list a native point cloud or 3D object type. CVAT includes a 3D task workspace for point cloud annotation in its standard editor, which makes it the usual choice of the two for lidar data.

Do Both Tools Support Segment Anything (SAM)?

Yes, in different ways. CVAT’s Community edition can run SAM as a serverless Nuclio function, and its paid plans include built-in SAM 2 and SAM 3 auto-labeling. Label Studio provides SAM and SAM 2 image examples for its ML backend, which you deploy and connect yourself, and the documented SAM 2 backend needs a GPU.

Which Has Better Quality Control In The Free Version?

CVAT’s free edition includes a manual review workflow with a validation stage and an issue tool. Label Studio’s comparison page lists reviewer workflows, agreement metrics and role-based access as paid features. Advanced QA such as ground truth jobs and annotator analytics sits in both vendors’ paid tiers.

Conclusion

CVAT and Label Studio are both mature, actively developed, permissively licensed tools with real companies behind them, and the CVAT vs Label Studio choice is less about quality than about fit. CVAT, built by a former Intel computer vision team, is stronger for images, video and 3D, and its free edition includes a basic review loop. Label Studio, from HumanSignal, covers far more data types through its configurable interface and is a natural fit for text, audio and LLM data work.

List the data types and annotation shapes you need, decide whether your team can maintain a self-hosted deployment, and check which QA features you would pay for. A short pilot on your own data in each tool usually settles the question.