WeVisOne-VL

Overview

WeVisOne-VL unifies object-level and pixel-level perception through structured next-token prediction. Coordinate tokens represent boxes and points, while a lightweight 9.6M-parameter decoder combines object-token states, boxes, and visual features to recover masks and continuous alpha mattes. The model covers detection, referring localization, counting, pointing, OCR, segmentation, and promptable matting, including absent-target responses.

Task	Input	Output
Object Detection	Category list	Labeled boxes
Referring Localization	Natural-language expression	Boxes
Pointing	Category or expression	Points
Counting with Boxes / Points	Category or expression	Boxes / points and count in the image description
OCR	Image	Recognized text as box labels
Text Grounding	Text to find	Text boxes
Text Segmentation	Category or expression	Polygons and binary mask PNGs
Text Matting	Category or expression	Polygons, alpha PNGs, and transparent RGBA cutouts
Box Matting	Positive boxes	Polygons, alpha PNGs, and transparent RGBA cutouts
Visual Prompting	Reference boxes	Similar-object boxes
Freeform	Instruction	Text and any predicted boxes, points, or masks

Installation

Follow the X-AnyLabeling-Server installation guide (English / 中文), including the wevisone_vl dependency extra.

Enable wevisone_vl in configs/models.yaml. Its configuration is configs/auto_labeling/wevisone_vl.yaml; the default checkpoint is tencent/WeVisOne-VL-2B. A local checkpoint directory is also accepted.

Usage

Open the AI panel, choose Remote-Server, connect to your server, and select WeVisOne-VL.

• Select a text task, enter your prompt, and click Send. Detection accepts comma-, period-, or newline-separated categories such as dog,plant. Referring and pointing prompts preserve the complete expression, for example the dog on the sofa.
• For counting, enter a category such as plants; the image description contains the number of returned annotations. For OCR, click Run without a prompt. For Text Grounding, enter the exact text to locate, such as TODAY.
• For Text Segmentation or Text Matting, enter dog or a referring expression. The contour fineness slider adjusts polygon simplification; mask_threshold in the server configuration controls the contour threshold.
• For Box Matting, draw positive boxes around the foreground and click the rectangle run button. For Visual Prompting, draw reference boxes around example objects and run. Use Finish to name the returned objects; Clear removes prompts.
• Freeform accepts instructions such as Describe this image.. The image description shows natural-language output with coordinate tokens and markup removed.
• Enable Preserve existing annotations to retain previous shapes. Text tasks support text-prompt batch processing, OCR supports ordinary batch processing, and reference-box tasks disable batch processing.

After matting, the upper-right result thumbnail shows the transparent foreground. Click it to open a larger preview. Multiple foreground instances are combined in preview.png; individual alpha maps and cutouts are also retained. The preview follows the current image, model, and task during the session.

Segmentation and matting exports are saved beside the source image in a new <image_stem>_outputs_* directory for every request. The application log records the directory. Binary masks preserve holes that exterior polygon contours cannot represent; alpha PNGs retain continuous opacity at 8-bit precision, and cutouts retain the original RGB pixels with that alpha channel. Unsaved input images use a temporary output directory.

Examples

Object Detection

Referring Localization

Pointing

Counting with Boxes

Counting with Points

OCR

Text Grounding

Text Segmentation

Text Matting

Box Matting

Visual Prompting

Freeform

Notes

The integration uses the Transformers backend. Model predictions may miss objects, misread text, or produce inaccurate counts; inspect the annotations before accepting them. No confidence or IoU controls are exposed because the model generates structured tokens rather than calibrated detection scores.