Tempered Guided Diffusion

TGD separates posterior inference in clean space from the conditional reconstruction module. TGD defines its annealed targets over clean signals rather than noisy diffusion states. Tempered Guided Diffusion uses interacting particles instead of independent trajectories to allocate computation during sampling. TGD treats noisy diffusion variables as auxiliary states for proposing clean reconstructions and moving particles between stages. TGD keeps the same outer scheme regardless of the conditional reconstruction module used.