Latent Space Arithmetic is the practice of performing algebraic operations (addition, subtraction, averaging) on latent vectors of a generative model to achieve compositional semantic editing, based on the discovery that well-structured latent spaces encode semantic concepts as consistent vector directions that can be combined through simple arithmetic. The classic example is the analogy: vector("king") - vector("man") + vector("woman") ≈ vector("queen"), which extends to visual attributes in generative models.

Why Latent Space Arithmetic Matters in AI/ML: Latent space arithmetic reveals that generative models learn compositional semantic structure where complex concepts decompose into additive vector components, enabling intuitive attribute transfer and compositional editing through simple vector operations.

• Concept vectors — Semantic attributes are encoded as directions in latent space: the "glasses" vector v_glasses can be computed by averaging latent codes of faces with glasses minus the average of faces without glasses, creating a transferable attribute direction • Attribute transfer — Adding a concept vector to any latent code transfers that attribute: z_with_glasses = z_face + v_glasses; subtracting removes it: z_without_glasses = z_face - v_glasses; this works because well-disentangled spaces encode attributes as approximately linear, independent directions • Analogy completion — Visual analogies follow the same pattern as word embeddings: z(man with glasses) - z(man without glasses) + z(woman without glasses) ≈ z(woman with glasses), demonstrating that the model has learned to separate identity from attribute • Multi-attribute editing — Multiple concept vectors can be combined additively: z_edited = z + α₁·v_smile + α₂·v_young + α₃·v_glasses, enabling simultaneous control over multiple independent attributes with separate scaling factors • Limitations — Arithmetic assumes attributes are linearly encoded and independent; in practice, attributes are often entangled (changing "age" may change "hair color"), and the linear assumption breaks down at large magnitudes

Latent space arithmetic is the most intuitive demonstration that generative models learn compositional semantic structure, enabling attribute transfer, analogy completion, and multi-attribute editing through simple vector addition and subtraction that reveals the linear, disentangled organization of knowledge within learned latent representations.