Improving blendshape performance for crowds with GPU and GPGPU techniques

For real-time applications, blendshape animations are usually calculated on the CPU, which are slow to animate, and are therefore generally limited to only the closest level of detail for a small number of characters in a scene. In this paper, we present a GPU based blendshape animation technique. By storing the blendshape model (including animations) on the GPU, we are able to attain significant speed improvements over CPU-based animation. We also find that by using compute shaders to decouple rendering and animation we can improve performance when rendering a crowd animation. Further gains are also made possible by using a smaller subset of blendshape expressions, at the cost of expressiveness. However, the quality impact can be minimised by selecting this subset carefully. We discuss a number of potential metrics to automate this selection.