FreeNeRF: Improving Few-Shot Neural Rendering with Free Frequency Regularization

Turning the left to the right by adding one line of code: pos enc[int(t/T * L)+3:]=0 Figure 1. Example novel view synthesis results from sparse inputs. The only difference between NeRF (left) and FreeNeRF (right) is the use of our frequency regularization, which can be implemented as few as, approximately, one line of code (bottom, where t and T denote the current training iteration and regularization duration, respectively; L is the length of the input positional encoding).