X-AnyLabeling is a desktop annotation tool with an AI inference engine built in, aimed at the unglamorous but decisive part of every computer vision project: producing labeled data. Released under GPL-3.0 and maintained by CVHub520, it has accumulated close to 10,000 GitHub stars since 2023 and remains under unusually active development — the changelog records new features landing nearly every week through 2026. Its pitch is straightforward: rather than drawing every box and polygon by hand, let foundation models produce the first pass and spend human time on correction instead of creation. ## What It Is The tool descends from the well-known labelme lineage but has grown into something considerably broader. It runs on Linux, Windows, and macOS, requires Python 3.11 or newer, and completed a full PyQt5-to-PyQt6 refactor in March 2026. Where most annotation tools handle one or two label types, X-AnyLabeling covers classification, detection, segmentation, captioning, rotated boxes, tracking, pose estimation, OCR, visual question answering, grounding, and document parsing in a single application, across both images and video. The supported annotation geometries are equally wide: polygons, rectangles, cuboids, rotated boxes, quadrilaterals, circles, lines, line strips, and points, plus dedicated modes for text detection, text recognition, and key information extraction. That range is what lets a team standardize on one tool rather than stitching together several. ## AI-Assisted Labeling The integrated model library is the core of the value proposition. Segment Anything provides promptable segmentation, and as of April 2026 the project added client-side local ONNX support for SAM 3 text-grounded segmentation — meaning a user can describe a target in words and get a mask without a network round trip. Grounding DINO and the newer LocateAnything support open-vocabulary grounding, pointing, and scene text detection, so novel object classes can be labeled without training a detector first. YOLO models cover conventional detection, PP-OCRv6 and PaddleOCR handle multilingual text detection, recognition, and document parsing, GeCo2 performs object counting, SCRFD 10G KPS does face detection with five-point landmarks, and TrackTrack provides multi-object tracking across video frames. Inference runs through flexible backends including ONNX Runtime, TensorRT, and OpenCV DNN, with TensorRT support for YOLO models added in April 2026 for teams that need throughput. One-click inference applies a model across every image in the current task, and a companion project, X-AnyLabeling-Server, adds remote inference for setups where the annotation machine is not the GPU machine. ## Workflow Maturity What distinguishes a research demo from a production tool is the surrounding workflow, and this is where the recent changelog is telling. The 2026 releases added an annotation check status workflow for reviewing labeled images, per-shape locking to prevent accidental geometry changes, interactive group selection with safe copy and paste, intent-aware selection that prioritizes nearby vertices when shapes overlap, an eraser for removing polygon vertices, a drag rotation handle for rotated boxes, and a collapsible panel for adjusting annotation opacity and image brightness. These are the fixes that emerge from people annotating thousands of images, not from a feature checklist. Import and export cover COCO, VOC, YOLO, DOTA, MOT, MASK, PPOCR, MMGD, VLM-R1, and ShareGPT — the last two notable because they target vision-language model training rather than classical CV formats. The UI is localized in English, Chinese, Japanese, and Korean. ## Pros and Cons The strengths are breadth without fragmentation: one application spanning image and video, classical detection through VQA and document parsing, with a deep library of current foundation models and export formats that reach both traditional CV pipelines and VLM fine-tuning. Local ONNX inference means sensitive datasets need not leave the machine, and the release cadence is genuinely fast. The trade-offs are real. GPL-3.0 is the most consequential — it is a copyleft license, and organizations that plan to modify and distribute derivative tooling need to understand the obligations, which is a different calculation than the Apache-2.0 and MIT terms common elsewhere in this space. The README carries an LGPL badge while the repository's actual LICENSE file is GPL-3.0, so teams should read the license itself rather than the badge. Beyond licensing, this is a desktop application rather than a collaborative web platform, so multi-annotator coordination, task assignment, and review queues are not built in the way managed services provide them. Running the larger models locally demands real hardware, and the sheer number of supported models and formats makes the initial learning curve steeper than a single-purpose tool. ## Outlook Annotation tooling is being reshaped by the fact that foundation models can now do a credible first pass on tasks that used to require full manual labeling. X-AnyLabeling's direction — adding SAM 3, open-vocabulary grounding, and VLM-oriented export formats — tracks that shift closely, and its expansion into VQA and document parsing suggests the maintainer sees multimodal dataset construction, not just bounding boxes, as where the work is heading. ## Conclusion X-AnyLabeling is a mature, actively maintained choice for teams that need to build vision datasets locally and want model assistance without sending data to a hosted service. It fits self-hosted CV teams and multimodal data engineers particularly well. Organizations that require collaborative review workflows or that cannot accept GPL-3.0 obligations should evaluate those two constraints before committing.