Letting Trajectories Spread: Quality-Preserving Control for Diverse Flow Matching

Flow-based text-to-image models follow deterministic trajectories, making it costly to explore diverse modes under limited sampling budgets. Existing approaches to improving diversity often rely on retraining or degrade image fidelity. To address this limitation, we present a training-free, inference-time control mechanism that makes the flow itself diversity-aware. Our core insight is to encourage diversity through guidance that is geometrically decoupled from the model’s quality-seeking direction. Our method simultaneously encourages lateral spread among trajectories via a feature-space objective and reintroduces uncertainty through a time-scheduled stochastic perturbation. Crucially, this perturbation is projected to be orthogonal to the generation flow, a geometric constraint that allows it to boost variation without degrading image details or prompt fidelity. Theoretically, we show that this design monotonically increases a volume surrogate while approximately preserving the marginal distribution, providing a principled explanation for the robustness of generation quality. Empirically, across multiple text-to-image settings under fixed sampling budgets, our method consistently improves diversity metrics such as the Vendi Score and Brisque over strong baselines, while upholding image quality and alignment.

Modern text-to-image systems can create realistic pictures from a prompt, but asking for several images often returns variations that look surprisingly similar. This limits their usefulness for users who want to explore different designs, scenes, or creative ideas. We introduce OSCAR, a method that helps a flow-based image generator explore a wider range of possibilities at generation time, without retraining the model. The idea is to make different generation paths spread apart from one another, while only pushing them in directions that do not disrupt the model’s normal path toward a high-quality, prompt-matching image. In this way, OSCAR acts like a gentle diversity guide: it encourages variety, but avoids the random changes that can create artifacts or ignore the prompt. Across several text-to-image settings, our experiments show that OSCAR produces more diverse image sets while preserving visual quality and alignment with the user’s text prompt.