Level 0: The Model

Taste Vectors

The geometry layer projects aesthetic understanding into a compact vector space, then uses that shared space to learn lightweight per-user taste manifolds.

Taste Projection In Geometry Layer

We use contrastive training to align the embeddings of images with aesthetic language.

Projection

Per-token hidden states project into 128-D multi-vector embeddings.

Global Tasks

A final pooled token is used for global tasks that need a compact representation.

Similarity

Retrieval uses ColBERT-style MaxSim geometry for fine-grained matching.

Per User Taste Manifolds

We can then build a lightweight model for every user that expresses their aesthetic likes and dislikes.

  • Operate only in 128-D space
  • Model both attraction and aversion
  • No retraining of the base model
  • Support personalized ranking, taste-aware discovery, and dislike diagnosis

Dislike Diagnosis

The same user manifold can answer the question: "what breaks the vibe?"