腾讯微信视觉开源 WeVisOne-VL:一个小模型兼顾检测、计数、OCR 与抠图
微信视觉团队开源基于 Qwen3-VL 的 WeVisOne-VL(2B/4B,Apache 2.0),用单一视觉语言模型覆盖开放词汇检测、指代定位、计数、OCR、分割与 alpha 抠图,并强调目标不存在时返回 None。
腾讯微信视觉团队开源 WeVisOne-VL,提供 2B 与 4B 两个体量,协议 Apache 2.0,底座为 Qwen3-VL。设计目标是:不要为「找物体、数数量、读文字、抠出来」各挂一套专用模型,而用一个视觉语言模型统一处理开放词汇检测、指代定位、计数、指点、OCR、指代分割与 alpha matting。
技术路径上,框与点以坐标 token 输出,再由约 950 万参数的像素解码器生成掩码与全分辨率 mattes。作者强调一个实用细节:当询问的目标不在图中时,模型会返回 None——对流水线来说,「诚实说没有」往往比硬猜一个框更安全。模型卡写明当前以图像为主,精细边缘仍受分辨率与遮挡影响;榜单成绩为作者自报,需用自有样本复核。
对做合成、电商抠图、文档理解或视觉 Agent 工具链的团队,可先拿十几张自有照片(含故意不含目标的负例)测检测与 matting 诚实率,再决定是否接入。细节以官方模型卡与仓库为准。
以上整理仅供行业观察,具体以官方文档与仓库为准。
以上整理仅供行业观察,具体以官方文档与仓库为准。
WeVisOne-VL open-sourced
How many separate vision models does it take to find, count, read and cut out objects in a photo? Tencent's WeChat Vision team is betting on one.
They open-sourced WeVisOne-VL in 2B and 4B sizes under Apache 2.0, built on Qwen3-VL. A single model handles open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation and alpha matting. Boxes and points come out as coordinate tokens, and a small 9.5M-parameter pixel decoder turns them into masks and full-resolution mattes.
The detail I'd test first: it returns None when the thing you asked for isn't in the image. A model that reliably says "not there" is much safer to wire into a pipeline. The benchmark wins are author-reported, and the model card says it's images only for now, with fine edges still sensitive to resolution and occlusion.
If you composite AI images, the matting alone is worth a look. Grab a dozen of your own photos, include a few where the target is missing, and count how often it answers honestly.
#ComputerVision
Link in Comments