---
format: "aidr-story-markdown/v1"
id: "6b3bc6cd2e55ab03f93a5982655975bf15d23f11fd1c2da29bb5c6d74f8b782b"
canonical_url: "https://aidr.today/6b3bc6cd?lang=en"
title: "FlowBudget: Training-Free Sampling-Budget Estimation for Flow Matching Models"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-16T04:00:00.000Z"
category: "Research"
topics: ["inference"]
source_urls: ["https://arxiv.org/abs/2610.20636"]
summary: "FlowBudget is a training-free estimator that determines the sufficient numerical resolution for sampling from continuous flow-based generators like flow matching models. It uses simulation-free interpolation probes to normalize the leading one-step discretization error by motion, estimates the coefficient from these probes, and maps a numerical tolerance to a step budget using the solver order. This addresses the problem of selecting sampling budgets through expensive FID-versus-NFE sweeps."
---

# FlowBudget: Training\-Free Sampling\-Budget Estimation for Flow Matching Models

> [Open the canonical story](<https://aidr.today/6b3bc6cd?lang=en>)

**Published:** 2026-10-16T04:00:00.000Z
**Category:** Research
**Topics:** inference

## Summary

FlowBudget is a training\-free estimator that determines the sufficient numerical resolution for sampling from continuous flow\-based generators like flow matching models\. It uses simulation\-free interpolation probes to normalize the leading one\-step discretization error by motion, estimates the coefficient from these probes, and maps a numerical tolerance to a step budget using the solver order\. This addresses the problem of selecting sampling budgets through expensive FID\-versus\-NFE sweeps\.

## Sources

- [Story source](<https://arxiv.org/abs/2610.20636>)