What is WeVisOne-VL?

WeVisOne-VL is a model released by Tencent: Qwen3-VL fine-tuned specifically for perception tasks.

It handles object detection, segmentation, counting, and OCR with a single model.

Florence-2 is a similar model, but the base model has come a very long way, so it naturally handles complex prompts better.

Downloading the Models

📂ComfyUI/
└── 📂models/
    └── 📂text_encoders/
        ├── wevisone_vl_2b_int8_convrot.safetensors
        └── wevisone_vl_4b_int8_convrot.safetensors

WeVisOne-VL Detect

Switch task to choose what it does.

segmentation

Describe the target in text, and it outputs masks and BBOXes.

wevisone_vl_segmentation.json
  • Turn on individual_masks and, when there are several targets, each one comes out as its own mask.

detection / referring

Basically segmentation without the mask output.

Strictly speaking, detection and referring use different prompts internally, but in my tests the results were almost the same. Since segmentation already outputs BBOXes too, I don't think you'll need these very often.

point

Instead of BBOXes, it gives the positions of objects as coordinates.

count_box / count_point

Counts the target objects.

wevisone_vl_count_box.json

The count comes out as a sentence like Total number: 6., which is a little awkward to use.

This workflow pulls out just the number with a regular expression in Extract Text, then turns it into an INT with Convert Number.

ocr

Extracts the text in an image.

wevisone_vl_ocr.json

It outputs the coordinates and content of each word as text.

Unfortunately, it doesn't seem to handle Japanese well.

ocr_grounding

The reverse of ocr: write the text you're looking for, and it outputs a BBOX showing where it is in the image.

wevisone_vl_ocr_grounding.json

visual_prompt

Finds objects that match the one you boxed, within the same image.

It doesn't work from a separate reference image, so it may be hard to put to good use.

wevisone_vl_visual_prompt.json

freeform

The tasks above each shape the input and output for their job. In this mode, whatever you write in query is passed as the prompt as is.

wevisone_vl_freeform.json

Write something like Describe this image. and you can use it as a regular VLM to describe images, too. That said, it was built for perception tasks, so for this kind of use, a regular Qwen3-VL would be a better fit.

Samples