Flow-based generative models trained with simulation-free approaches have become strong contenders to diffusion models in various tasks. However, training-free guidance methods for controlled generation for flow models are not as well developed compared to diffusion. In this work, we propose Tweedie Moment Projected Flows (TMPF) model for zero-shot controlled generation using a pre-trained flow model. We demonstrate the effectiveness of the proposed method on linear inverse problems on both synthetic and real-world data such as high-resolution images.
This repository is based the Rectified Flow implementation (Liu et al. 2023).
Some qualitative results are shown below on the AFHQ-cats dataset (Choi et al. 2019):
