Currently, mesh has a concept of LoD, where as the object is at a distance, the number of rendered tris are reduced to a lower LoD model. I am proposing a similar concept with image textures where for instance at close distances, the mesh's 1024 textures are used, but pull back a short distance, a system resized texture of 512 is used, and further back, a 256 is used. This should reduce the vRAM usage, at the cost of more network usage to download the smaller textures as required.