Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching

Researchers propose learning condition-dependent source distributions for flow matching in generative models, demonstrating that optimizing the source distribution—rather than defaulting to standard Gaussian—significantly improves text-to-image generation performance. The approach achieves up to 3x faster convergence in FID scores while addressing stability challenges through variance regularization and directional alignment techniques.

Flow matching has gained traction as a computationally efficient alternative to diffusion models for generative tasks, yet the field has largely inherited the Gaussian source distribution assumption without critical examination. This research challenges that convention by treating source distribution selection as an optimization problem itself, particularly when rich conditioning signals like text prompts are available. The researchers identified concrete failure modes—distributional collapse and training instability—that emerge when naively incorporating conditioning information, then developed principled solutions through variance regularization and source-target alignment mechanisms.

The work builds on growing recognition within generative AI that architectural choices inherited from earlier methods may not be optimal at scale. As text-to-image systems have matured, practitioners increasingly question fundamental assumptions, from training objectives to sampling strategies. Condition-dependent source distributions represent a natural extension of this questioning, acknowledging that different prompts may benefit from different starting distributions rather than a one-size-fits-all approach.

The practical implications are substantial. A 3x improvement in FID convergence speed directly reduces computational costs and training time, making advanced generative models more accessible. The stability improvements matter for production systems where reliability is critical. For developers and researchers, these findings suggest that source distribution design deserves investigation alongside other optimization targets. The analysis of how target representation spaces interact with structured sources provides a framework for future improvements. This represents incremental but meaningful progress in generative model efficiency—not a breakthrough, but a systematic refinement with tangible benefits for practitioners deploying these systems at scale.

• →Optimizing source distributions under flow matching improves FID convergence speed by up to 3x compared to standard Gaussian approaches
• →Condition-dependent source distributions risk collapse and instability without proper variance regularization and directional alignment constraints
• →Different conditioning signals benefit from different source distributions rather than using a universal Gaussian baseline
• →Target representation space choice significantly impacts effectiveness of structured source distributions in flow matching
• →Research demonstrates that inherited assumptions from diffusion models warrant critical reevaluation in flow matching architectures