Expanding Flow Maps: Generative Flows for Variable-Dimension and Variable-Length Outputs

Forum topic · 小凯 · 2026-07-26

Summary

A Chinese tech forum post introduces the arXiv paper 2507.20479, "Expanding Flow Maps" by Sophia Tang and Pranam Chatterjee. Flow-based generative models achieve fast, controllable generation over continuous and discrete state spaces, but existing parameterizations are constrained to fixed dimensions or fixed sequence lengths. The paper introduces Expanding Generative Flows (EFlows), which define flows between distributions of increasing dimensionality along an expanding interpolant that grows the state by augmenting it with conditional noise. Building on this, the authors propose Expanding Flow Maps (EFMs), a new class of flow maps that distill the expanding interpolant into efficient few-step generative models. Each EFM factors the mapping between any two timesteps into two learnable operations: an expand operator that augments the state space with new coordinates or tokens conditioned on the current state, and a transport map that advances the expanded state along the interpolant. Composing these operations yields a single map that jointly expands and denoises the state, recovering existing fixed-canvas flows and flow maps as special cases where the expand operator is the identity. The framework extends to discrete simplices, enabling variable-size graph generation and variable-length sequence generation, establishing EFlows and EFMs as a principled framework for settings where output size is itself a learnable, controllable degree of freedom.

Overview

This post shares the paper "Expanding Flow Maps" by Sophia Tang and Pranam Chatterjee (ML), published on arXiv as 2507.20479.

Original Abstract

Flow-based generative models have enabled remarkable progress in fast and controllable generation across continuous and discrete state spaces, yet existing parameterizations are constrained to fixed dimensions or fixed sequence lengths. Here, we introduce Expanding Generative Flows (EFlows), which define flows between distributions of increasing dimensionality along an expanding interpolant that grows the state by augmenting it with conditional noise. Building on this construction, we propose Expanding Flow Maps (EFMs), a new class of flow maps that distill the expanding interpolant into efficient few-step generative models. Each EFM factors the map between any two timesteps into two learnable operations: an expand operator, which augments the state space with new coordinates or tokens conditioned on the current state, and a transport map, which advances the expanded state forward along the interpolant. Composing these operations yields a single map that jointly expands and denoises the state, recovering existing fixed-canvas flows and flow maps as special cases when the expand operator is the identity. The framework further extends to discrete simplices, enabling variable-size graph generation and variable-length sequence generation.

Key Takeaways

Problem: Flow-based generative models are limited to fixed dimensions or sequence lengths, restricting their use when output size should vary.
EFlows: Define flows between distributions of increasing dimensionality via an expanding interpolant that grows the state with conditional noise augmentation.
EFMs: Distill the expanding interpolant into efficient few-step generative models by factoring each two-timestep map into a learnable expand operator and a transport map.
Generality: Composed expansion + transport jointly expands and denoises the state; existing fixed-canvas flows/flow maps are recovered as the special case of an identity expand operator.
Applications: Extends to discrete simplices, enabling variable-size graph generation and variable-length sequence generation.

Paper Link

arXiv: 2507.20479

*Auto-collected on 2026-07-26*