Neural rendering

Rendering is the process of generating photo-realistic images from 2D or 3D computer models. The term neural rendering has recently emerged to describe rendering using a neural network. In traditional 3D rendering, we will need to first create a 3D model with a polygon mesh that describes the object's shape, color, and texture. Then, the lighting and camera position will be set and render the view into a 2D image.

There has been an ongoing research on 3D object generation, but it is still not able to generate satisfying results. We can take advantage of the advancement of GANs by projecting part of the 3D objects into 2D space. We then use GANs to enhance the image in 2D space, for example, to generate a realistic texture using style transfer before projecting that back into the 3D model. The top diagram in the following figure shows the general pipeline of this approach:

Figure 10.12 – Two common frameworks for neural rendering...