Tencent's WeVisOne-VL does detection, segmentation and OCR in one pass

Tencent's AI lab unveiled WeVisOne-VL, a vision-language model that emits bounding boxes, masks and OCR text as a single token stream. On one RTX 4090 a frame takes 48 ms versus 152 ms for a YOLO-v8, SAM and PaddleOCR stack, with GPU memory down from 12 GB to 6 GB.

• 48 ms latency versus 152 ms for the three-model pipeline on RTX 4090
• GPU memory cut from 12 GB to 6 GB, FLOPs down 58%
• 44.2 AP on COCO-Panoptic, within 0.3% of best task-specific ensembles
• Trained on 1.8B image-text pairs over 28 days on 64 A100 GPUs