AligNeRF: High-Fidelity Neural Radiance Fields via Alignment-Aware Training

Our AligNeRF Ground Truth (Warped) w/ Alignment-aware Training w/o Alignment-aware Training Estimated Optical Flow

Figure 1. Left: Images rendered by mip-NeRF 360 [2] (w/o alignment-aware training) and our AligNeRF (w/ alignment-aware training), with the same amount of training time for both methods. Middle: Magnified cropped regions. Here, the ground truth has been aligned with the mip-NeRF 360 rendering, as described in Sec. 4.3. Right: Visualization of misalignment. This is the estimated optical flow between the mip-NeRF 360 rendering and the original ground truth images.