The most visible ways that nvidia and ati are tackling the LOD problem is on-GPU tessellation. This means using a mathematically smooth surface (such as beziers or nurbs), and approximating it with polygons at the appropriate level for each frame. This means that if your lettuce is far away, it gets few polys (or culled). If it's 20% of your screen, it gets lots more. You can avoid popping artifacts because it's the same surface, just approximated differently, and it's not too tricky to get approximations that blend smoothly into each other.
If you need more detail that doesn't need to be animated, toss on a displacement map at relatively low memory cost.
True volumetric modeling works best for things that have lots of branching and void spaces, like dandelions.
Displaced surfaces work for lots of interesting things. See Mudbox and ZBrush galleries:
Mudbox http://usa.autodesk.com/adsk/servlet/pc/index?siteID=123112&...
Apparently there are drawbacks to this rendering approach, although this company has claimed to have solved them. I'd be interested to hear from anybody here what experience they have with this type of rendering.