Skip to main content
WeVisOne-VL is a compact vision-language model that answers in shapes instead of prose. Give it an image and a short text query, and it returns boxes, points, or a per-instance alpha matte in one pass: no annotation, no per-object prompting. In ComfyUI it runs as the WeVisOne-VL Detect node. The template below wraps it in a subgraph and previews the matte, the boxes drawn over the source image, and foreground and background cutouts, so the matte drops straight into a background-removal, masking, or compositing chain. The weights are the Apache-2.0 Tencent releases, repackaged for ComfyUI in Comfy-Org/WeVisOne-VL. Support landed in ComfyUI #17632 and the template needs ComfyUI 0.45.0 or later.
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you can’t find them in the template, your ComfyUI may be outdated.If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

Image Segmentation: WeVisOne-VL

Type what you want cut out in query and run. With the default segmentation task and the query human, the graph returns one combined matte of the people it finds, plus their boxes. Turn on individual_masks on the node for a separate matte per person. WeVisOne-VL segmentation workflow preview with matte, boxes and cutouts

Download Workflow

Download JSON or search “Image Segmentation: WeVisOne-VL” in Template Library
Input material

utility_image_segment_wevisone_vl_input.png

Load into the LoadImage node

Tasks

The task widget rewrites your query into the instruction the checkpoint was trained on. Each task takes its query in a slightly different form: Keep the query short. A full sentence makes segmentation ramble, and a phrase that names the thing works better than a description of it.

Models

The workflow loads the 2B INT8 text encoder. The other three files are drop-in swaps for clip_name in the CLIPLoader node (stable_diffusion type): the 4B pair is heavier and slower, and the BF16 files trade size for precision.

2B INT8 (this workflow)

wevisone_vl_2b_int8_convrot.safetensors → ComfyUI/models/text_encoders/

2B BF16

wevisone_vl_2b_bf16.safetensors → ComfyUI/models/text_encoders/

4B INT8

wevisone_vl_4b_int8_convrot.safetensors → ComfyUI/models/text_encoders/

4B BF16

wevisone_vl_4b_bf16.safetensors → ComfyUI/models/text_encoders/

More control on the node

  • individual_masks: on, one mask per box; off, one union mask per frame
  • max_length: answer length cap, 4096 tokens by default and up to 32768
  • Outputs: masks, bboxes, points (JSON with x, y, label and frame), and text (the raw answer)