2025-10-29
Starting from an implicit neural network that generates a red cube based on the orientation of the camera, we can retrain it with conditioned one-hot labels for three different shapes: a green capsule, a red cube, and a yellow sphere.
We toggle the shape input using the 'U', 'I', and 'O' keys to pass in the relevant one-hot vector. The rendered shape instantly changes:
Because our neural network is generating the image per-pixel (its input consists of the pixel position and the one-hot class label), the result has out-of-distribution properties: we can generate shapes that exist between the shapes.
The below video shows this in action. The shape interpolation starts a few seconds in.
These results are similar to EngramNCA models. They show that a single Neural Cellular Automata (NCA) network trained on three shape types can generate both shapes at once. This is due to the distributed, message-passing based behavior in NCAs. This bias allows for this output to exist, unlike traditional 'one-shot' image generation approaches.