Inference-Time Compute Scaling For Flow Matching

2025-10-20

AI Summary

Flow Matching (FM) models face a fundamental limitation: their inference-time computational budget cannot be scaled to improve generation quality without compromising efficiency. This work addresses this challenge by proposing a computation-scaling method for FM inference that preserves the linear interpolation property of standard sampling. The core innovation is a variance-preserving dynamic interpolation strategy, which maintains straight-line sample trajectories while enabling flexible increases in either integration steps or model evaluations during inference. To our knowledge, this is the first approach to achieve scalable inference computation for FM without sacrificing its inherent efficiency. We further extend the method to unconditional protein structure generation—a demanding scientific application. Experiments demonstrate monotonic improvements in both image and protein structure generation quality with increasing inference compute, validating the method’s cross-domain applicability, scalability, and practical utility.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Diffusion Models for Vision

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Allocating extra computation at inference time has recently improved sample quality in large language models and diffusion-based image generation. In parallel, Flow Matching (FM) has gained traction in language, vision, and scientific domains, but inference-time scaling methods for it remain under-explored. Concurrently, Kim et al., 2025 approach this problem but replace the linear interpolant with a non-linear variance-preserving (VP) interpolant at inference, sacrificing FM's efficient and straight sampling. Additionally, inference-time compute scaling for flow matching has only been applied to visual tasks, like image generation. We introduce novel inference-time scaling procedures for FM that preserve the linear interpolant during sampling. Evaluations of our method on image generation, and for the first time (to the best of our knowledge), unconditional protein generation, show that I) sample quality consistently improves as inference compute increases, and II) flow matching inference-time scaling can be applied to scientific domains.

Problem

• Developing inference-time compute scaling for Flow Matching
• Preserving linear interpolant during sampling unlike prior methods
• Extending flow matching scaling to scientific domains like proteins

Innovation

• Preserves linear interpolant during sampling
• Applies compute scaling to protein generation
• Maintains efficient straight sampling in flow matching