CAR-Flow：條件感知重參數化對齊源域與目標域以實現更優的流匹配

摘要

條件生成建模旨在從包含數據-條件對的樣本中學習條件數據分佈。為此，基於擴散和流的方法已取得了令人信服的成果。這些方法使用一個學習到的（流）模型，將初始的標準高斯噪聲（忽略條件）轉移到條件數據分佈。因此，模型需要同時學習質量傳輸和條件注入。為了減輕模型的負擔，我們提出了條件感知重參數化流匹配（CAR-Flow）——一種輕量級的學習偏移，用於對源分佈、目標分佈或兩者進行條件化。通過重新定位這些分佈，CAR-Flow縮短了模型必須學習的概率路徑，從而實現了更快的實際訓練。在低維合成數據上，我們可視化並量化了CAR的效果。在高維自然圖像數據（ImageNet-256）上，為SiT-XL/2配備CAR-Flow將FID從2.07降低到1.68，同時引入的額外參數不到0.6%。

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. 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.

CAR-Flow：條件感知重參數化對齊源域與目標域以實現更優的流匹配

CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matching

摘要

Support