Jag: Direct Box Prediction for Efficient Visual Grounding
Jag: Direct Box Prediction for Efficient Visual Grounding
Jag 是一個 0.8B 視覺定位模型,能在單次前向傳遞中預測邊界框,無須自回歸生成座標。它在五個測試分組上均優於基礎模型,整體 Acc@0.5 與 LocateAnything 相若;單次請求基準測試中,速度快 3.62×,峯值 GPU 記憶體用量少 73.88%。Jag 已公開模型權重、訓練與評估程式碼,並提供互動示範。
We introduce Jag, a 0.8B visual grounding model that predicts bounding boxes in a single forward pass, without autoregressive coordinate generation.
Jag improves over its base model on all five test splits and achieves overall Acc@0.5 comparable to LocateAnything. In our single-request benchmark, it is 3.62× faster than LocateAnything and uses 73.88% less peak GPU memory.
We release the model weights, training and evaluation code, and an interactive demo.
🌐 Project · 💻 Code · 🤗 Model · 🎮 Demo
Try the demo and let us know what you think. Questions and feedback are welcome!
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co