WeVisOneDetect - ComfyUI Built-in Node Documentation - ComfyUI

WeVisOne-VL Detect finds objects in an image from a text prompt and returns boxes, alpha mattes, points, or the model’s own answer, depending on the task you pick. It covers open-vocabulary detection and segmentation, referring expressions, pointing, counting and OCR grounding, and runs one prompt per image in the batch.

Inputs

Parameter	Description	Data Type	Required	Range
`clip`	The WeVisOne-VL text encoder that loads the checkpoint.	CLIP	Yes	-
`image`	The image or batch of images to search.	IMAGE	Yes	-
`task`	What the node should return. Segmentation returns boxes plus alpha mattes; detection, referring, ocr and ocr_grounding return boxes; point and count_point return points; count_box counts with boxes; visual_prompt finds objects similar to the reference boxes in `bboxes`; freeform sends the query unchanged, so the reply is normally text (default: “segmentation”).	COMBO	Yes	`"segmentation"` `"detection"` `"referring"` `"point"` `"count_box"` `"count_point"` `"ocr"` `"ocr_grounding"` `"visual_prompt"` `"freeform"`
`query`	What to find, in plain language (for example “the red mug”). `detection` and `point` read one label per line, `count_box` and `count_point` count the subject you name, `ocr_grounding` looks for this exact text, `visual_prompt` uses `bboxes` instead, and `freeform` sends it as the whole prompt. `ocr` ignores it.	STRING	Yes	-
`bboxes`	Reference boxes that `visual_prompt` uses as examples of the object to find.	BOUNDING_BOX	No	-
`individual_masks`	Return one matte per detected box from the `bboxes` output, frame by frame, instead of one combined matte per frame (default: false).	BOOLEAN	No	true false
`max_length`	Maximum number of tokens the model may generate for its answer (default: 4096).	INT	Yes	1 to 32768

Note: Alpha mattes are filled when the model answers with object references, which the segmentation task asks for. Other tasks normally return boxes, points or text with an empty matte, but a freeform prompt that asks for object references can fill them too.

Outputs

Output Name	Description	Data Type
`masks`	Alpha mattes of the matched objects, one per frame by default or one per box with `individual_masks`.	MASK
`bboxes`	Boxes found in each image, stored as `x`, `y`, `width`, `height` and `label`.	BOUNDING_BOX
`points`	Points as JSON in pixels, one object per point with `x`, `y`, `label` and `frame`, for example `[{"x": 128, "y": 96, "label": "cup", "frame": 0}]`. Filled by the `point` and `count_point` tasks.	STRING
`text`	The model’s raw answers, joined with newlines; an answer can itself span several lines.	STRING

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 0cef1dfcab790e145ace51401a84cdc10282835330cda5a3770f5b20a32d475e