学术研究★★★arXiv · 2026-07-23
Expanding Flow Maps: Dynamic Dimensionality Generation Models
Flow-based generative models have made significant progress in generating data quickly and controllably across various state spaces, but they are limited to fixed dimensions or sequence lengths. This paper introduces Expanding Generative Flows (EFlows) and proposes Expanding Flow Maps (EFMs) to dynamically increase state dimensions through conditional noise.
📌 Key points
- EFlows dynamically expand state dimensions by adding conditional noise
- EFMs transform the expanding interpolant into efficient flow mappings
- This approach overcomes the fixed dimension constraints of traditional flow mode
本页为 gitzw.com 基于公开来源的 AI 中文解读，非原文转载。