WeVisOne-VL now runs in ComfyUI core: one node turns a text prompt into boxes, masks, alpha mattes, points, counts or OCR
Tencent WeChat Vision's WeVisOne-VL, an Apache-2.0 vision model in 2B and 4B sizes, runs natively in ComfyUI since 9 November 2026, when kijai's pull request was merged. A single Detect node takes an image and a phrase and returns boxes, per-instance alpha mattes, masks, points, counts or OCR boxes, without SAM or custom nodes.
What it does
The model is a Qwen3-VL fine-tune with a 9.6 million parameter mask decoder; the node's tasks cover segmentation, detection, referring, point, counting, OCR, OCR grounding, visual prompts and free-form questions. For cutouts you wire the masks and image into InvertMask and JoinImageWithAlpha, and asking for something that is not in the frame returns None instead of a made-up box.
Worth knowing
It works on single images only, not video, so it is not a tracking node, and fine alpha edges can still drift on heavy occlusion or transparent subjects. The support is on ComfyUI master and was not yet in the v0.44.2 tag; the Comfy-Org repack loads through a CLIPLoader (type stable_diffusion) and comes in four files from 2.9 GB (2B int8) to 9.7 GB (4B bf16).