SubFlow: Sub-mode Conditioned Flow Matching for Diverse One-Step Generation

2026-04-14

AI Summary

Existing single-step flow matching generative models, trained with mean squared error, often suffer from submode averaging, which suppresses low-density yet valid distributional modes and compromises generation diversity. To address this, this work proposes a submode conditioning mechanism: first, semantic clustering partitions each data class into intrinsic submodes; then, flow matching is performed conditioned on submode indices, encouraging each conditional subdistribution to be approximately unimodal. This approach requires no architectural modifications and can be seamlessly integrated as a plug-and-play solution to eliminate averaging artifacts. Evaluated on ImageNet-256, the method significantly improves generation diversity (measured by Recall) while maintaining excellent image quality (FID), demonstrating its general effectiveness across single-step frameworks such as MeanFlow and Shortcut.

Technology Category

• Machine Learning: Multimodal Learning
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Generation

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Flow matching has emerged as a powerful generative framework, with recent few-step methods achieving remarkable inference acceleration. However, we identify a critical yet overlooked limitation: these models suffer from severe diversity degradation, concentrating samples on dominant modes while neglecting rare but valid variations of the target distribution. We trace this degradation to averaging distortion: when trained with MSE objectives, class-conditional flows learn a frequency-weighted mean over intra-class sub-modes, causing the model to over-represent high-density modes while systematically neglecting low-density ones. To address this, we propose SubFlow, Sub-mode Conditioned Flow Matching, which eliminates averaging distortion by decomposing each class into fine-grained sub-modes via semantic clustering and conditioning the flow on sub-mode indices. Each conditioned sub-distribution is approximately unimodal, so the learned flow accurately targets individual modes with no averaging distortion, restoring full mode coverage in a single inference step. Crucially, SubFlow is entirely plug-and-play: it integrates seamlessly into existing one-step models such as MeanFlow and Shortcut Models without any architectural modifications. Extensive experiments on ImageNet-256 demonstrate that SubFlow yields substantial gains in generation diversity (Recall) while maintaining competitive image quality (FID), confirming its broad applicability across different one-step generation frameworks. Project page: https://yexionglin.github.io/subflow.

Problem

• flow matching
• generation diversity
• mode collapse
• one-step generation
• sub-mode representation

Innovation

• Sub-mode Conditioning
• Flow Matching
• One-Step Generation
• Diversity Preservation
• Averaging Distortion