通过条件重参数化缩短生成路径，提升流匹配效率
CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matching
- 引入条件感知重参数化，动态调整源与目标分布位置
- ImageNet-256上FID从2.07降至1.68，参数增益低于0.6%
- 轻量级设计，适合高维图像生成任务快速部署
条件生成建模旨在从包含数据-条件对的样本中学习条件数据分布。扩散与基于流的方法在此任务上表现优异。这些方法利用学习到的（流）模型，将忽略条件的初始标准高斯噪声传输至条件数据分布。因此，模型需同时学习质量传输与条件注入。为减轻模型负担，我们提出条件感知重参数化流匹配（CAR-Flow）——一种轻量级、可学习的偏移机制，用于对源分布、目标分布或两者进行条件性调整。通过重新定位这些分布，CAR-Flow缩短了模型需学习的概率路径，从而在实践中实现更快训练。在低维合成数据上，我们可视化并量化了CAR-Flow的效果。在更高维的自然图像数据（ImageNet-256）上，使用CAR-Flow的SiT-XL/2模型将FID从2.07降至1.68，且额外参数少于0.6%。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。