Flow Where You Want – Guidance for Flow Models (drscotthawley.github.io)

A new tutorial has been released demonstrating how to implement inference-time controls for pretrained flow-based generative models, enabling them to perform tasks outside their original training scope. By utilizing classifier guidance and inpainting techniques, the authors show how to steer an unconditional flow model—which initially generates random MNIST digits—toward generating specific digits and filling in missing data. This guidance is accomplished by applying velocity corrections during the sampling process, allowing for targeted outcomes in both latent and pixel spaces. This development is significant for the AI/ML community as it introduces flexible methods to enhance flow models without the need for retraining, thereby broadening their applicability in real-world scenarios. The tutorial includes practical applications of these methods, emphasizing the ease of integrating guidance techniques into existing flow models like FLUX and Stable Audio. By exploring the intricacies of velocity adjustments and utilizing gradient descent for corrections, researchers can better manipulate flow-generated outputs, making it a valuable resource for those working with generative models.