Building Next-Gen Multimodal Foundation Models for General-Purpose Assistants

LLaVA is an open-source project, collaborating with research community to advance the state-of-the-art in AI. LLaVA represents the first end-to-end trained large multimodal model (LMM) that achieves impressive chat capabilities mimicking spirits of the multimodal GPT-4. The LLaVA family continues growing to support more modalities, capabilities, applications and beyond.

LLaVA represents a cost-efficient approach to building general-purpose multimodal assistant. It is a novel end-to-end trained large multimodal model that combines a vision encoder and Vicuna for general-purpose visual and language understanding, achieving impressive chat capabilities mimicking spirits of the multimodal GPT-4 and setting a new state-of-the-art accuracy on Science QA.

evolution

Recent development