X-AnyLabeling

X-AnyLabeling is a desktop annotation workspace for researchers and teams preparing text, image, video, point cloud, and multimodal training datasets. Models generate initial labels, and reviewers correct them in the same workspace before exporting the data. It runs on Windows, Linux, and macOS under the GPL-3.0 open-source license.

Annotation, review, and export happen locally. Model inference can run on your machine through ONNX Runtime, TensorRT, OpenCV DNN, or PyTorch, or use remote services such as SGLang, vLLM, and TGI. Supported model families include YOLO, SAM, DINO, Qwen, and PPOCR, and you can integrate custom models.

The workspace covers detection, segmentation, pose estimation, tracking, OCR, captioning, and visual question answering. Its drawing tools include polygons, rotated boxes, masks, points, and cuboids, so it can represent object outlines, positions, and 3D annotations within one application. Batch prediction lets models label datasets for human review.

Document tools use PaddleOCR to parse layouts, tables, formulas, and text into editable annotations. Video tools label frame-accurate segments and export clips or frame sequences. A vision-language chatbot works alongside the current image and saves approved responses as training data.

Dataset import and export support formats including COCO, VOC, YOLO, DOTA, MOT, PPOCR, and ShareGPT. The interface is available in English, Simplified Chinese, Japanese, and Korean.

Install from the official package, a GUI installer or source. Some advanced integrations, including SAM video tracking and selected model workflows, require a Git-clone installation. Model weights have their own licenses, and remote inference sends the relevant data to the configured service.