Implicit Dynamical Flow Fusion (IDFF) for Generative Modeling

TL;DR

IDFF introduces a new vector field with momentum for generative modeling, significantly reducing network evaluations needed for high-quality sample generation across images and time-series data.

Contribution

It proposes Implicit Dynamical Flow Fusion (IDFF), a novel method that accelerates sample generation by enabling longer steps while maintaining fidelity, reducing NFEs by tenfold.

Findings

Achieves comparable likelihood and quality to existing models with fewer NFEs

Reduces sampling time significantly without loss of fidelity

Demonstrates versatility across image and time-series datasets

Abstract

Conditional Flow Matching (CFM) models can generate high-quality samples from a non-informative prior, but they can be slow, often needing hundreds of network evaluations (NFE). To address this, we propose Implicit Dynamical Flow Fusion (IDFF); IDFF learns a new vector field with an additional momentum term that enables taking longer steps during sample generation while maintaining the fidelity of the generated distribution. Consequently, IDFFs reduce the NFEs by a factor of ten (relative to CFMs) without sacrificing sample quality, enabling rapid sampling and efficient handling of image and time-series data generation tasks. We evaluate IDFF on standard benchmarks such as CIFAR-10 and CelebA for image generation, where we achieve likelihood and quality performance comparable to CFMs and diffusion-based models with fewer NFEs. IDFF also shows superior performance on time-series datasets modeling, including molecular simulation and sea surface temperature (SST) datasets, highlighting its versatility and effectiveness across different domains.\href{https://github.com/MrRezaeiUofT/IDFF}{Github Repository}