Conservative Flows: A New Paradigm of Generative Models

TL;DR

This paper introduces a novel generative modeling paradigm called Conservative Flows, which uses data-supported states and stochastic dynamics to generate data, improving over traditional noise-based methods.

Contribution

It proposes a new paradigm for generative models utilizing data-supported states and develops probability-preserving sampling methods compatible with pretrained flow models.

Findings

The framework successfully generates data on synthetic and real datasets.

The proposed samplers outperform original generation procedures.

Validated on Swiss-roll, ImageNet-256, and Oxford Flowers-102.

Abstract

Modern generative modeling is dominated by transport from a noise prior to data. We propose an alternative paradigm in which generation is performed by a discrete stochastic dynamics that leaves the data distribution invariant, initialized from data-supported states rather than from noise. The framework can utilize any pretrained flow model. We develop two probability-preserving sampling mechanisms, a corrected Langevin dynamics with a Metropolis adjustment and a predictor-corrector flow, that operate directly on existing checkpoints. We validate the framework on a synthetic Swiss-roll target, ImageNet-256 and Oxford Flowers-102, where our samplers consistently improve over the original generation procedures.