Comment by sunnygo

great! have you tested how well it return coordinates of objects/text? Ive tried with generic LLMs like Gemini/Qwen/Gemma and they all are unstable with coordinates around text, better when using visual grounding though