Go with the Flow (Map): Fast Generation Meets Inference-time Scaling

Diffusion and flow models have recently driven major advances in generative modeling, yet their multi-step sampling remains slow and can falter on difficult generation tasks. In this talk, we turn to flow maps, which directly learn the noise-to-data mapping and connect arbitrary noise levels in a single step. This enables both highly efficient few-step sampling and flexible multi-step sampling. We first introduce Align Your Flow, a state-of-the-art flow-map framework that scales to fast text-to-image generation. We then show that flow maps are not only powerful for fast generation but also excel at steering diffusion model sampling toward challenging user-specified rewards in a test-time compute scaling setting. Our framework, Flow Map Trajectory Tilting, uses a fast and accurate flow-map look-ahead—rather than the standard denoiser look-ahead—to obtain precise reward estimates and effectively guide slow stochastic generative processes. This enables efficient guidance by complex reward functions, including those defined by vision-language models. Finally, we highlight inference-time optimization applications in protein design, making the case for flow map-based generative modeling and test-time scaling beyond image generation. In particular, we discuss Complexa, which leverages inference-time search to design protein binders for challenging targets.