Note The training data used in the paper

Note data with bounding box coordinates

osunlp/UGround-V1-2B

Image-Text-to-Text • 2B • Updated Feb 16 • 2.12k • 8

Note Based on Qwen2-VL-2B-Instruct

osunlp/UGround-V1-7B

Image-Text-to-Text • 8B • Updated Apr 16 • 3.7k • 17

Note Based on Qwen2-VL-7B-Instruct

osunlp/UGround-V1-72B

Image-Text-to-Text • 73B • Updated Jan 23 • 58 • 4

Note Based on Qwen2-VL-72B-Instruct. Full training without LoRA.

osunlp/UGround-V1-72B-Preview

Image-Text-to-Text • 73B • Updated Jan 12 • 10 • 2

Note Based on Qwen2-VL-72B-Instruct. Trained with LoRA.

osunlp/UGround

Image-Text-to-Text • 7B • Updated Apr 16 • 117 • 23

Note The initial model. Based on the modified LLaVA arch (CLIP + Vicuna-7B) describe in the paper

Note Low-cost, scalable and effective data synthesis pipeline for GUI visaul grounding; SOTA GUI visual grounding model UGround; purely vision-only (modular) GUI agent framework SeeAct-V; first time demonstrating SOTA performance of vision-only GUI agents.

Note Paused. Will open a new one for Qwen2-VL-based UGround

Note Paused. Trying to figure out how to accelerate the inference.