What to notice: The lines of the polygons correspond roughly to surface shapes like the collarbone, the shape of the breast, the muscles of the neck, or the loosely-defined ridge between the ribs and the soft midriff. But there is no way to embed any knowledge of the underlying structure of the woman in the polygon cage. There is no concept of how, if she were to flex her shoulder muscle, its volume would change. There is no understanding of how, if her breasts were to fill or empty of milk, they would change shape.
Traditional rigging techniques allow multiple vertex geometries depending on how the bones move (Blender does not yet have this capability but Maya, for example, does). The artist then moves the bones to various extremes, models the vertexes how they should look there, then the software morphs the vertexes from one geometry to the next as the bone is rotated from one extreme to the next.
All this could be avoided if only the polygon cage had knowledge of the muscle, bone, fat, tendon, and ligament structures that made it up. Here I believe metaballs hold the key to better organic modelling. Follows an illustration of a dummy modelled up with metaball modelling implicit surfaces:
What to notice (I will speak of weaknesses now): When the arm gets too close to the body, it begins to melt into it, as though they were one structure. The weakness of metaballs currently is that they melt into all other metaballs of the same class. In order to correctly model organics, a metaball must be able to melt into only one other metaball, or only a couple others. This way, the forearm could melt into the elbow, which melts into the upper arm.
Even this is still a little weak. An organic being isn't made up of structures that melt into each other. Simply speaking, it's a skin that is shrink-wrapped onto structures underneath it, but I tend to think the abstraction of metaballs doesn't need to go that far to present a convincing model.
What to notice that is a strength of metaballs: The knee structure, the leg muscles, the midriff, the breasts, the shoulders, the neck, and the head all are clearly made up of masses. We are building up the organism from lumps of clay, rather than creating an infinitely thin shell that looks like a person.
I can quickly grab the "stomache" metaball, make it larger, and suddenly the organism looks fatter. A quick adjustment to its shape, and she looks pregnant. I can quickly grab the shoulder muscles, make them larger, and she appears to have stronger shoulders. All such operations on a polygon-modelled mesh are far more difficult to do convincingly.
Even if metaballs are modified as I propose, there are some problems to solve. The polygon model is textured, and as the model is moved into different poses, it is easily defined how the texture will stretch and deform to follow the new pose, whereas such a definition isn't easily made with metaballs, since the surface of the resulting model doesn't have a static morphology or even topology of vertexes.
Another commonly-mentioned shortcoming of metaball modelling is that the artist does not have precise control of the number of polygons dedicated to various regions of the model. Given computers are still quite limited in their power as of 2005, this is perhaps a criticism that would require a bit of thought.
In the end, however, I feel that if metaballs doesn't provide the key to excellent organic modelling, there will regardless be some other modelling method that defines underlying muscle and tissue structure that will supplant polygon modelling (and indeed, if you follow Zbrush's modelling tools, you might notice they have a concept called the Zsphere which looks like a method for defining a metaball structure that gives rise to a polygon model).
Philo Vivero gives away all his models (in .blend.gz format) at http://faemalia.net/Blender -- note that there is, like this page, nudity there.