YOLO-World

YOLO-World is the Tencent AI Lab project that finally gave the YOLO family open-vocabulary abilities. It adds a Re-parameterizable Vision-Language Path Aggregation Network (RepVL-PAN) and a region-text contrastive loss to a standard YOLO backbone, pretrained on large-scale image-text and detection data. You describe what you want in plain text and the model detects it, zero-shot, with no retraining. At inference the text embeddings can be precomputed and folded into the network, so most of the open-vocabulary overhead disappears and it runs like a normal YOLO.

The CVPR 2024 paper reports 35.4 AP on LVIS at 52 FPS on a V100, beating previous open-vocabulary methods on both accuracy and speed, and fine-tuned variants transfer well to downstream detection and open-vocabulary instance segmentation. The research code is GPL-3.0, which is fine for research but copyleft for commercial products, and Ultralytics also ships its own YOLO-World variants under AGPL-3.0, so check which license your deployment actually falls under.

Author-reported (CVPR 2024): 35.4 AP on LVIS at 52.0 FPS on V100; zero-shot YOLO-World-L 26.8 LVIS AP / 44.9 COCO AP