CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matching
This work addresses efficiency and performance in conditional generative modeling for applications like image generation, though it appears incremental as it builds on existing flow-based methods.
The paper tackles the challenge of conditional generative modeling by proposing CAR-Flow, a lightweight method that conditions source and target distributions to shorten probability paths, resulting in improved performance such as reducing FID from 2.07 to 1.68 on ImageNet-256 with minimal parameter overhead.
Conditional generative modeling aims to learn a conditional data distribution from samples containing data-condition pairs. For this, diffusion and flow-based methods have attained compelling results. These methods use a learned (flow) model to transport an initial standard Gaussian noise that ignores the condition to the conditional data distribution. The model is hence required to learn both mass transport and conditional injection. To ease the demand on the model, we propose Condition-Aware Reparameterization for Flow Matching (CAR-Flow) -- a lightweight, learned shift that conditions the source, the target, or both distributions. By relocating these distributions, CAR-Flow shortens the probability path the model must learn, leading to faster training in practice. On low-dimensional synthetic data, we visualize and quantify the effects of CAR-Flow. On higher-dimensional natural image data (ImageNet-256), equipping SiT-XL/2 with CAR-Flow reduces FID from 2.07 to 1.68, while introducing less than 0.6% additional parameters.