LabelBlend AI Studio documentation

Installation

LabelBlend AI Studio is a Windows desktop application. Install it from the Microsoft Store for automatic updates, or use the release published on GitHub if you prefer a direct download.

The installer does not bundle AI model weights. Models are downloaded separately from inside the application so that the install size stays small and you only fetch the checkpoints you actually use.

• ·Install from the Microsoft Store, or download a release from GitHub.
• ·Launch the application and choose a theme under the Theme menu.
• ·Open the model manager and download the checkpoints you need.
• ·Select an image folder to create your first project.

Getting Started

The left rail switches between workspaces: Prompt, Video, Mask, Edit, Blend and View. Each workspace targets one stage of dataset creation and shares your classes and folders with the others.

Start in Edit if you want to annotate manually, or in Mask if you plan to use SAM2 assistance from the beginning.

Projects

A project is an image folder plus its labels and class definitions. Use Open Project and Save Project in the Edit workspace to resume a session later.

Keeping images, labels and classes.txt in a consistent structure allows the viewer to auto-detect the dataset.

Annotation

Rectangle mode produces bounding boxes. Polygon and pen modes produce outlines suitable for segmentation. Select mode lets you pick an existing annotation to move, reshape, reclassify or delete.

Add classes in the Object Classes panel and assign each one a colour before you start labelling.

SAM2

SAM2 provides click-driven mask proposals. Choose a checkpoint and its matching config file, connect, then click on an object to receive a proposal.

Larger checkpoints generally produce better masks but require more VRAM. If your GPU is limited, start with a smaller checkpoint.

• ·Select a checkpoint (for example a hiera small or base plus variant).
• ·Select the matching config file if it is not detected automatically.
• ·Press Connect and wait for the status to report a connected model.
• ·Click an object, choose a class, then confirm the annotation.

Text Prompting

Prompt-driven labelling accepts lowercase phrases separated by dots, for example: car. person. traffic light.

Box threshold and text threshold control how confident a detection must be before it is proposed. Lower values propose more regions and require more review.

Video Tracking

Import a video, choose whether to resize frames, and load frames with either a RAM limit or a frame-skip interval. Memory indicators show current VRAM and RAM usage.

Add the objects to be tracked, then run tracking forward or backward across the loaded frames. Free RAM and reload frames if you change the resize settings.

Synthetic Data

Add background images and object cutouts in the Blend workspace, then set minimum and maximum object size, rotation range and how many images to produce.

Annotations for every placed object are generated automatically. Box and mask datasets can be exported from the same panel.

Dataset Validation

The View workspace loads a dataset folder and renders annotations over each image. Switch between box view and polygon view, adjust line width and fill opacity, and toggle class labels.

Correct any issues in the Edit workspace and re-export before training.

Export

Set output size and quality, optionally enable a train/validation split, and choose the destination folder. Box datasets and mask datasets are exported separately.

Vertices epsilon controls how aggressively polygon outlines are simplified when exporting mask datasets.

Model Downloads

Large AI models do not need to be bundled inside the installer. Instead, the application downloads the checkpoints you select and stores them locally.

Workflow: install LabelBlend → open the model manager → choose an AI model → download → use it locally. Models remain on disk between sessions, so the download only happens once per checkpoint.

Checkpoints are stored under your Windows user application data folder. Plan for several gigabytes of free storage if you intend to keep multiple checkpoints.

Settings

Theme selection lives in the Theme menu, with a light productivity theme and a dark engineering theme.

Output profiles, thresholds, CPU thread counts and memory limits are set per workspace and persist between sessions.

Keyboard Shortcuts

Shortcuts are shown in the application status bar while you work.

• ·A — previous image · D — next image
• ·S — select · R — rectangle · P — polygon
• ·Enter — confirm · C — cancel · Del — delete annotation
• ·Space (hold) — grab and pan the canvas
• ·Esc — cancel the current AI operation

Troubleshooting

Common issues and where to look first.

• ·Model will not connect: confirm the checkpoint finished downloading and that the config file matches the checkpoint.
• ·Out of VRAM during tracking: enable low VRAM mode, reduce frame size, or increase the frame skip value.
• ·Frames fail to load: check the video codec and try extracting frames first from the Blend workspace.
• ·Viewer shows no annotations: verify the folder structure contains images/, labels/ and classes.txt, or set the paths manually.
• ·Export folder is empty: confirm a destination folder is selected and that at least one confirmed annotation exists.

FAQ

Short answers to the questions asked most often.

• ·What is LabelBlend? An AI-powered computer vision data preparation and dataset creation workspace for Windows.
• ·Who should use it? AI engineers, researchers, students, startups and independent developers preparing detection or segmentation datasets.
• ·Is it free? A free tier covers basic annotation, basic augmentation and basic export. Pro covers the AI workflows.
• ·What is included in Pro? AI-assisted annotation, SAM2 workflows, text-driven labelling, video tracking, video-to-dataset, synthetic data generation, batch workflows and advanced annotation.
• ·Does it work offline? Annotation, editing, composition and export run locally. Downloading models and opening documentation require an internet connection.
• ·Does it use SAM2? Yes, SAM2 checkpoints are used for assisted segmentation and video tracking.
• ·Can it process video? Yes — import a video, load frames and track objects across them.
• ·Can it create synthetic datasets? Yes, by composing object cutouts onto background images with automatic annotations.
• ·Can I edit AI-generated annotations? Yes. Every proposal is editable in the annotation editor before export.
• ·Can I export training datasets? Yes — box datasets and mask datasets, with an optional train/validation split.
• ·Where are AI models stored? In your Windows user application data folder, downloaded on demand.
• ·How do I install it? From the Microsoft Store, or from a release published on GitHub.
• ·Which operating systems are supported? Windows. Other platforms are not supported today.
• ·How can I get support? Open an issue on GitHub or use the contact page.