How to implement skeletal animation for Signed Distance Field models?

Question

When animating a 3D model, we want the surface of the model to remain continuous (such that, for instance, the limbs do not seem to slide on the torso, but stay attached at their position).

In a triangle-based mesh, this is easy since the edges of mesh preserve where a limb gets attached to torso. In fact, preserving the volumetric properties becomes the tricky part, to ensure that faces don't end up intersecting each other, etc.

However, in an SDF, while the volumetric properties get preserved automatically, the surface properties do not. Instead of the character looking like a creature with stretchable skin than holds their muscles and bones inside, it looks like a cloud whose various volumes can slide in, out or around upon movement.

What are the ways that skeletal animation can be applied to Signed Distance Fields, if any? If there are none, what alternative to skeletal animation might be used to achieve same effect?

Answer 1

I've thought about this a lot myself and think the following problems make certain kinds of animation more difficult with sdfs than polygons:

• smoothing between sdf primitives has a lot to do with the nice features of sdf modelling, but this means that a model might start smoothing with itself undesirably as you animate it. consider a hand and fingers, all made with a series of smoothed primitives. As fingers get closer together, like a folded hand, a fist, etc, the fingers will all blend in with each other to create a new whole rather than individual fingers.

• the old "squash and stretch" traditional animation exercise. With polygons, it is trivial to elongate a ball, for example, and have it stretch easily in one direction while narrowing in another dimension to maintain volume. If you have an sdf sphere as your ball, this can be more complicated. perhaps you use non-uniform scaling to stretch the ball.. that is already undesirable as most sdf renderers stick to uniform scale (though non-uniform can be done) for mathematical reasons. the other option then is to use multiple smoothed spheres to create your stretched oval-like ball, but now you need to carefully tweak their placement and alter their sizes so the volume seems visually preserved. it's just not so useful; very hard to animate something like with all the changes you have to make at every frame compared to a basic polygonal skeleton.

• arbitrary scenes of many complex objects are hard to optimize with sdfs. With polygons, you have tried-and-true bounding hierarchies that can be automatically created at the triangle level and bounced off to a renderer, and many GPUs now handle these natively, some even create the bounding chains for you. With sdfs, if you want a rendering application to automatically handle a large number of primitive operations, the methods for doing this are not particularly obvious. demo scenes can hard-code in the bounding volumes as an optimization, but if you want to work with arbitrarily complex scenes, you can't expect someone to program this directly for every scene, especially as they start animating (nor would this be a reasonable thing to expect artists to do either). there are plenty of ways to optimize sdfs but each has its own drawbacks and can be exceptionally computationally expensive as a pre-render step in some cases (like voxel discretization) hence why marching cubes back to polygons is a clearer path to rendering efficiency, especially once sdfs start animating around.

I'd be curious what other things you've considered as ways to solve this problem?