I found something definitive: (I didn’t realize they were quite that small though)
At very small scales, even low-genus objects run into visual changes caused by over-decimation. Nanite overcomes this by pre-rendering 12×12 pixel imposter images from 144 different view directions in advance and using the image that most closely aligns with the current view direction if the object’s on-screen bounding box is less than 12 pixels in size. That solution requires memory for the set of imposters and can result in popping, so the Nanite devs have expressed a desire for something better. It also doesn’t do much to help with grass, leaves, hair, screens, blinds, and other large objects made from the cumulative visual effect of many small pieces.
https://cs418.cs.illinois.edu/website/text/nanite.html#case-study-nanites-approach-to-lod