Latent Process Generator Matching Framework Introduced
A new theoretical framework called latent process generator matching has been introduced, generalizing existing generator matching and flow-matching methods. The framework treats the observed generative state as a deterministic image of a tractable Markov process, allowing learning of a generator for a stochastic process on the image space with matching one-time marginal distributions. This subsumes discrete latent process results from the literature.
Key facts
- Framework introduced: latent process generator matching
- Generalizes generator matching and flow-matching methods
- Observed state X_t = Φ(Y_t) where Y_t is a tractable Markov process
- Learns generator of stochastic process on image space
- One-time marginal distributions match projected process
- Subsumes discrete latent process results
- Published on arXiv with ID 2605.20547
- Announce type: cross
Entities
Institutions
- arXiv