T-Rex Label

T-Rex Label is a browser-based annotation tool for building computer-vision datasets, built around a simple premise: you should have to label an object once, not five hundred times.

One prompt, batch labelling. Draw a bounding box around a single instance — one cow in a field, one bolt on a production line — and the model finds every other instance in that image, reports a detection count, and then carries the same prompt across the rest of your images automatically as you page through them. The workspace shows a confidence slider so you can tune the threshold before committing, and an "Auto-Detect this category in all images" toggle that applies the prompt dataset-wide.

Open-set models, no training step. The pre-annotation pipeline runs DINO-X and Grounding DINO 1.6; interactive annotation runs T-Rex2 and successor models. All are open-set detectors, which is the point — they recognise categories that were never in their training data, so there is no fine-tuning round and no chicken-and-egg problem where you need a labelled dataset to get a model that helps you label your dataset.

Visual prompts, not just text prompts. Most zero-shot labelling tools take a text description. T-Rex2 takes a visual prompt — the box you drew — which handles the cases text struggles with: a specific defect type, an unusual part, a category with no clean English name, or dense scenes where a text prompt over-matches.

Masks as well as boxes. Alongside bounding boxes, an AI Mask tool auto-segments every target after you select one, with positive and negative pens for correcting edges and a brush for larger areas. Mask annotation is part of the free interactive annotation service.

Practical details. It runs entirely in the browser with nothing to install or deploy, supports up to 100 categories per project, shows per-image annotation stats, and exports to the formats the ecosystem already uses, with documented paths into Kaggle, Hugging Face and Label Studio.

Who it is for. ML engineers and data teams assembling detection or segmentation datasets — manufacturing defect inspection, retail shelf auditing, agriculture, traffic and logistics analysis — and researchers who need a labelled set quickly without standing up annotation infrastructure.

Built by Visincept (formerly IDEA CVR), the team behind the T-Rex2 and DINO-X detection models.