Why do we keep sampling from the same distribution the model was trained on? We rethink this old paradigm by introducing Feynman-Kac Correctors (FKCs) – a flexible framework for controlling the distribution of samples at inference time in diffusion models! Without re-training or fine-tuning (just with the raw power of math), we demonstrate how to sample from: 1. product of densities (e.g. for conditioning generations on several things) 2. annealed density (invaluable property for diffusion-based samplers) 3. geometric-average of densities (to bridge the gap between classifier-free guidance and the common intuition) 4. reward-tilted density (easy to get within our framework, but we focus on other cases that don't require an external reward function) @martoskreto prepared beautiful animations for you in the thread below, or you can go directly to check - Paper: arxiv.org/abs/2503.02819 - Code: github.com/martaskrt/fkc-dif… - Educational example: colab.research.google.com/dr…