What Is FlyLabelImg? Offline AI-Assisted Image Annotation for Windows

Ask five teams how they label images and you will get five stacks: LabelImg for quick boxes, Label Studio on a server for the "real" project, a spreadsheet for class management, and a pile of Python scripts to convert formats. FlyLabelImg exists because that pile is annoying. It is a single Windows app for bounding-box and polygon annotation that adds the thing manual tools never had: a local vision-language model that pre-annotates your images, plus a review workflow that makes accepting or rejecting those drafts faster than drawing them yourself.

Why another annotation tool in 2026?

The classic desktop labelers (LabelImg, LabelMe) are solid but strictly manual — every box costs the same three clicks, whether it is the first image or the ten-thousandth. The web suites (CVAT, Label Studio) add AI and collaboration, but they want a server, accounts and your images uploaded somewhere.

The gap in the middle is where FlyLabelImg sits: AI-assisted, fully offline, zero-setup desktop annotation for individuals and small teams — images never leave the machine, which matters for medical, industrial and internal data.

Every step, one consistent app

Step	What happens	Notes in FlyLabelImg
Open a folder	Recursive image scan with thumbnails	All / unlabeled / pending / done filters; sidecar files stay next to the images
AI pre-annotate	Local model labels the current image or the whole folder	Confidence threshold, class prompts, concurrency capped by your license seats
Review drafts	AI results arrive as pending boxes, never written silently	Enter confirms, Esc rejects, A/D walks between pending items
Re-run AI	Drafts overlapping confirmed boxes are deduplicated	Same-class results with IoU > 0.7 are dropped automatically
Quality check	Rule-based and model-based label audits	Out-of-bounds, tiny, duplicate, unknown-class, format conflicts; optional model re-inference
Export	Pascal VOC XML or YOLO txt sidecars	Switchable per project — no conversion scripts

The point is not tidiness. It is that the moment you switch from drawing a polygon to judging twenty AI drafts, the canvas, the class colors and the keyboard shortcuts all behave identically.

What is actually inside

• Local AI pre-annotation — a bundled vision-language model (Qwen3.5-VL via llama.cpp) runs on your machine; the tray app starts and supervises it. Point it at any OpenAI-compatible endpoint instead, if you prefer.
• Review-first pending workflow — drafts, confirmations and rejections are tracked per image; a filter shows what is left to judge.
• Overlap dedup — re-annotate an old dataset with a better model without duplicating confirmed boxes.
• Rectangle and polygon tools — on a smooth, zoomable canvas with per-class colors and live counts.
• Annotation checker — out-of-bounds, tiny, duplicate and unknown-class boxes, plus format conflicts, in one pass.
• VOC XML and YOLO txt sidecars — switchable per project; polygon classes export to YOLO-seg style lines.
• Keyboard-first — W to draw, arrow keys to flip images, Enter/Esc to judge drafts. Hands stay on the keyboard.
• 10 interface languages — switchable at runtime.
• Dark, focused UI — designed for long labeling sessions.

What FlyLabelImg is not

Being explicit here saves you a download:

• It is Windows 10/11 (64-bit) only. There is no macOS or Linux build.
• It is not a team platform: no project server, no roles, no cloud sync, no dataset marketplace.
• It is not a model-training tool: it prepares datasets; training happens in your own framework.
• It has no telemetry, no ads and no bundling, and it needs no account.

Trying it

The download is a 30-day trial with every feature enabled — no crippled mode, no nag beyond the registration reminder. Licences are one-time purchases: $39 for one year or $99 for five years, updates included for the term.

FAQ

Is FlyLabelImg free?

There is a 30-day free trial with all features. After that, licences are one-time purchases: $39 for 1 year or $99 for 5 years.

Do my images get uploaded anywhere?

No. Annotation and AI inference run entirely on your machine. The only network traffic is the license verification, and only if you register. Details in the privacy policy.

How is it different from LabelImg?

FlyLabelImg adds local AI pre-annotation, a pending-review workflow, annotation checking and per-project VOC/YOLO switching in a modern dark UI. See the annotation tool comparison for the honest side-by-side.

Does the AI need a GPU or an internet connection?

The bundled model runs on a normal Windows PC; internet is not required at any point. For heavy daily use, a machine with a discrete GPU speeds inference up noticeably.