Neural Implicit Representations
Traditional rendering uses discrete data structures—like a 2D grid of pixels for an image, or a 3D grid of voxels for a volume. A Neural Implicit Representation discards the grid entirely. Instead, a neural network (a Multi-Layer Perceptron, or MLP) acts as a mathematical function that you can query at any continuous point in space.
You pass in a continuous spatial coordinate (x, y, z), and the network outputs the density and color at that exact microscopic point. This allows for mathematically infinite resolution without the massive memory overhead of 3D voxel grids.
The Problem (Spectral Bias): Neural networks are fundamentally biased toward learning smooth, low-frequency functions. If you feed raw, linear coordinates directly into a neural network, it struggles to unroll them into sharp edges, resulting in blurry, blobby outputs.
The A/B Comparison
To fix this spectral bias, we use Fourier Features (Positional Encoding). Instead of passing the raw coordinate x, we preprocess it into a high-dimensional array of expanding sine and cosine waves: [sin(x), cos(x), sin(2x), cos(2x), sin(4x)...].
- Left (Raw Coordinate MLP): Receives only raw, linear mathematical coordinates. Notice how it produces a smooth, featureless, low-detail blob. It cannot physically learn complex shapes.
- Right (Fourier Features): By expanding the coordinates into frequency bands, the network immediately "sees" the potential for sharp, high-frequency patterns and complex geometry.
The Positional Encoding Slider
The slider controls how many high-frequency sine/cosine bands are analytically fed into the network's input layer:
- 0 Bands: Identical to the broken raw coordinate model. Smooth and blurry.
- Medium Bands (4–6): The sweet spot for most NeRFs. It captures sharp edges and distinct shapes without adding chaotic noise.
- High Bands (10+): Captures microscopic, razor-sharp details. However, pushing this too high introduces high-frequency "ringing" artifacts and severe overfitting, where the network memorizes noise instead of the actual geometry.