t8star/WeVisOne-VL-Comfy

WeVisOne-VL for ComfyUI · T8

English | 简体中文

Original Tencent WeVisOne-VL checkpoints and Mask2Alpha assets, organized for ComfyUI-WeVisOne-VL-T8. T8 maintains the ComfyUI integration and this distribution; upstream model authors and licenses are preserved.

Files

Path	Contents
2B/	WeVisOne-VL 2B, tokenizer, processor and required pixel decoder
4B/	WeVisOne-VL 4B, tokenizer, processor and required pixel decoder
sam2.1_hiera_base_plus.pt	Official SAM2.1 Hiera Base+ checkpoint
mask2alpha_sam2_bplus.pt	Tencent Mask2Alpha refinement head
mask2alpha_head.onnx	Optional fixed-shape ONNX export of the refinement head
provenance.json, SHA256SUMS	Source revisions, file sizes and SHA-256 checksums

The pretrained weights are copied unchanged. TensorRT engines are built locally for the target GPU/runtime.

Use in ComfyUI

Install the node package and follow its setup instructions. The downloader selects only the requested model size and the shared Mask2Alpha checkpoints.

For manual download:

from huggingface_hub import snapshot_download

snapshot_download(
    "t8star/WeVisOne-VL-Comfy",
    local_dir="ComfyUI/models/wevisone",
    allow_patterns=["2B/*", "sam2.1_hiera_base_plus.pt", "mask2alpha_sam2_bplus.pt"],
)

Set WEVISONE_MODELS to that directory, or select the model path in the loader. Replace 2B/* with 4B/* for 4B. mask_decoder.pt is required for text-guided segmentation and matting.

Sources and license

• Tencent WeVisOne-VL · Mask2Alpha branch
• Official 2B · Official 4B · Meta SAM2

Apache-2.0. Original notices are retained in each model directory and LICENSES/. Pinned provenance is recorded in provenance.json.

T8 links

Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI pack · Hugging Face