Tsinghua KEG Lab and Zhipu AI jointly launched CogAgent, a large image understanding model

Bit News Tsinghua KEG Lab recently cooperated with Zhipu AI to jointly launch a new generation of image understanding large model CogAgent. Based on the previously launched CogVLM, the model uses visual modalities instead of text to provide a more comprehensive and direct perception of the GUI interface through a visual GUI agent for planning and decision-making. It is reported that CogAgent can accept 1120×1120 high-resolution image input, with visual question answering, visual positioning (Grounding), GUI Agent and other capabilities, in 9 classic image understanding lists (including VQAv2, STVQA, DocVQA, TextVQA, MM-VET, POPE, etc.) has achieved the first result in general ability.