Latent Space Navigator
Where Data Dreams: Visualise the Compressed Reality of Generative AI.
Latent Space Navigator
The Compressed Code (The Latent Vector)
Every complex piece of data—be it an image, a sentence, or an audio clip—can be boiled down to a simple, short list of numbers called the Latent Vector. This vector is the “compressed code” of the original data. This tool maps these latent vectors into a 3D coordinate system.
Nodes clustered together in this space represent items that are mathematically similar (e.g., all images of a car are near each other, but far from all images of a bird).
Smooth Interpolation (The Morph)
The key property of the Latent Space is its smoothness. By selecting two distant points (Source A and Target B) and moving a slider, you can observe interpolation—the process of creating new, never-before-seen data that smoothly bridges the gap between the source and the target.
As you move the slider, the tool traces a path through the 3D space, showing the creation of intermediate data forms (e.g., a “dog-cat” hybrid).
Disentanglement (The Controllable Feature)
The best latent spaces are disentangled, meaning one dimension (or axis) controls one specific feature (e.g., rotating the head, changing hair colour, or adjusting emotional tone).
This tool allows you to manipulate the space along its primary axes. Watch as moving along the X-axis changes the data in one specific way (like moving from light to dark), demonstrating the controllable power of the AI’s internal representation.
