if you allowed for some slippage in precision, how many images could you fit in that are recognizable from various angles, with ~acceptable~ resolution? Sounds fun.
Not exactly what you describe but this project [0] did it with cubes.
The cube is filled with splats. Each face reveals a different picture when viewed from a perpendicular angle
Imagine you are holding in your hand a small glassy sphere. Initially it looks like random dots inside, but as you play with it in your hands, you get glimpses of images of loved ones.
To implement this, select a bunch of photos of people, generate pseudo-3d models of their faces with your AI of choice. Position them around a sphere in a 3D modelling software, and generate a whole bunch of renderings all around the scene. Now, feed those renderings into gaussian splat software.
The resulting 3D gaussian splat should include relatively high fidelity when the viewer is aligned with the supplied iamges, but more stochastic nonsense when misaligned. This should contribute to the feeling of being able to 'find' images while rotating the sphere around.