Neural rendering · Experiment

Fun with NeRF

A playful experiment that turns four related images into a lenticular animation by changing how a neural radiance field samples images and rays.

PyTorchNeRFNGPReconstruction
Frames from a NeRF-based lenticular reconstruction
Lenticular reconstruction generated from a small collection of related images.

The idea

Instead of treating NeRF only as a tool for conventional novel-view synthesis, I wanted to see what would happen if the input images represented changing subjects or moments. The goal was an image that appears to shift as its virtual viewpoint changes—similar to a lenticular print.

How it works

The project builds on my simplified Instant NGP implementation. A custom data-loading path changes how the four input images and their rays are presented during training and testing.

  • Train from four images with roughly matching dimensions.
  • Control the effect through the data loader and generated camera extrinsics.
  • Render the learned representation as an animated GIF.

Why it matters

The experiment is small, but it makes an important idea tangible: the behavior of a learned scene representation depends as much on the structure of its observations as on the network itself. Changing the sampling process can turn a reconstruction pipeline into a creative rendering tool.