Implicit Dynamical Flow Fusion (IDFF) for Generative Modeling
This addresses the computational bottleneck in generative modeling for applications like image and time-series data generation, though it is incremental as it builds on existing CFM methods.
The paper tackles the slow sampling speed of Conditional Flow Matching (CFM) models by proposing Implicit Dynamical Flow Fusion (IDFF), which reduces network evaluations by a factor of ten while maintaining comparable sample quality on benchmarks like CIFAR-10 and CelebA.
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}