generate new data by learning to reverse a gradual noising process. Often we need samples that combine multiple models, like one trained on cats and one on dogs, or that maximize a reward. Naively mixing their outputs gives biased results.
The authors developed a method called Generalized Feynman-Kac Correctors. It tracks the probability of each model's output along the generation and corrects the combination on the fly. This works for any target function and also allows steering toward a reward, all without retraining.
On simple tests, the method closely matched the true distribution. On MNIST handwritten digits, it successfully combined concepts from different models. This lets users customize outputs by reusing existing models, saving time and resources.