Stereo Radiance Fields (SRF): Learning View Synthesis for Sparse Views of Novel Scenes

Figure 1. Our method can synthesize new views with a single network forward pass from 10 sparse and spread-out views of a novel scene. Here we synthesize a new view with (a) NeRF [34], which requires scene-specific training for 2 days; with our method (b) which produces the result instantaneously and (c) our improved result after fine-tuning our pre-trained model for 15 minutes on the 10 views.