ModelLLaVA-UHD v4
LLaVA-UHD v4
A multimodal vision encoding architecture in the LLaVA series. It introduces early visual token compression within the ViT, reducing subsequent computation by merging or discarding redundant shallow-layer tokens, and has been adopted by several on-device multimodal models.
Timeline (last 90 days)
Oct 12
On the vision side, MiniCPM-V 4.7 adopts the early compression inside the ViT from LLaVA-UHD v4.
Unverified50%