We build a Monte Carlo Tree over the diffusion denoising process that can be used for scalable, compute-efficient, inference‑time alignment of pretrained diffusion models to new reward functions