WeVisOne VL · T8
WeVisOne visual perception, Mask2Alpha, geometry and video tools for ComfyUI
ComfyUI WeVisOne-VL · T8
26 ComfyUI nodes for WeVisOne-VL 2B/4B, Mask2Alpha, and image/video masks. Supports detection, OCR, text-guided matting, interactive prompts, and SAM2 video propagation.
Models · Workflows · Validation · Issues
Features
- Vision: detection, referring expressions, counting, point grounding, OCR and text localization.
- Matting: text/box-to-alpha; refine masks with boxes, positive/negative clicks, or polygons.
- Visual editor: multi-object prompts, video keyframes, undo/redo, and JSON import/export.
- Video: chunked SAM2 tracking, scene reset, checkpoints, temporal stabilization, timestamps and audio export.
- Integration: standard
IMAGE/MASK, native boxes, kijai points/BBOX, Impact SEGS, ROI/tiles and mask utilities. - Acceleration: feature reuse, int8/NF4, ONNX/TensorRT head runtimes and optional WSL/Linux vLLM.
Install
Install WeVisOne VL · T8 in ComfyUI-Manager (Registry ID: wevisone-vl-t8, publisher: t8star), or clone into ComfyUI/custom_nodes:
git clone https://github.com/T8mars/Comfyui-WeVisOne-VL-T8.git
The Manager installs lightweight host dependencies. Prepare the isolated inference worker once, from this repository directory, using uv:
uv run --no-project --python 3.12 tools/setup_worker.py --models 2B
The setup installs CUDA PyTorch and model dependencies into this package's .venv, then downloads verified weights into .models. It leaves the host's Transformers stack separate. Use --models 2B 4B for both sizes, --acceleration for quantization/ONNX dependencies, or --device cpu for CPU setup. Restart ComfyUI and search WeVisOne. Workflows are in workflows/; load the .json files without .api in their names.
Manual installations also need python -m pip install -r requirements.txt in the ComfyUI environment. Advanced users can set WEVISONE_PYTHON to another prepared worker Python and WEVISONE_MODELS to another model directory. in_process requires the backend dependencies in that interpreter.
Start here
| Workflow | Purpose |
| --- | --- |
| 01_text_alpha | Text → soft alpha → transparent PNG |
| 02_bbox_refine | Box → Mask2Alpha |
| 03_video_track | Keyframe prompts → tracking/refinement → alpha video |
| 04_visual_picker | Click, draw boxes/polygons, then refine |
| 05_accelerated_refine | Visual prompts with a locally built TensorRT engine |
| 06_remote_vision | Recognition through a running vLLM service |
For the visual editor, run once to load the preview. Left-click adds a foreground point; right-click excludes background. Select auto in RefineAlpha. For OCR, connect LoadVLM → Vision, then choose ocr or ocr_grounding.
Mask convention: 1 = foreground/opaque. For ComfyUI LoadImage's transparency MASK, select transparent_area. Boxes use original-image, half-open pixel xyxy. Video frames retain their batch order.
Performance
On an RTX 4060 Ti 16 GB, repeated-image tests measured about 1.45× faster Mask2Alpha refinement with TRT FP16 and 3.88× faster recognition with vLLM. Including fresh SAM2 image encoding, TRT FP16 improved about 10%. Quantization primarily reduces VRAM; ONNX/compile results depend on the platform. See test conditions and setup and acceleration instructions.
Credits
Built on Tencent WeVisOne-VL, its Mask2Alpha branch, and Meta SAM2. Apache-2.0; upstream notices and model licenses are preserved. See THIRD_PARTY_NOTICES.md.
T8 links
Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI pack · Hugging Face