Skip to content

Generative JEPA

A focused subtopic within the JEPA documentation series: how to extend a Joint Embedding Predictive Architecture from a representation-learning model into a generative one, with the concrete target of perturbation response prediction on single-cell data.

Vanilla JEPA predicts embeddings, not data. It learns the semantic structure of how inputs relate — but it cannot emit a sample, define a likelihood, or quantify uncertainty on its own. This subtopic works through what it takes to close that gap, surveys the 2025–2026 literature that does so, and commits to a concrete architecture.

Documents

Doc Contents
architecture_spec.md The design space (four routes to a generative JEPA), the empirical caveat that drives the choice, and a concrete staged architecture — Generative Perturbation JEPA (GP-JEPA) — with module/loss/evaluation breakdown.

Why this is its own subtopic

The base series (00–05) treats JEPA as predictive and notes "use a generative wrapper" without specifying one. That hand-wave hides two separate problems — stochastic prediction and decoding back to data space — and the right way to solve them depends on recent work (Var-JEPA, Cell-JEPA, D-JEPA, V-JEPA 2 world models). This subtopic exists to treat that properly and to host follow-on research as it lands.

Relationship to the rest of the project

  • Builds directly on 04_jepa_perturbseq.md (the predictive baseline).
  • Reuses the count-aware decoders in src/genailab/model/ and the latent diffusion infrastructure in src/genailab/diffusion/.
  • Realizes the JEPA component of the three-stage pipeline described in PROJECT_OVERVIEW.md.