Training-Free Conditional Diffusion

The TGD framework unifies several existing training-free samplers under different schedule and weighting choices. Training-free conditional diffusion reuses an unconditional diffusion prior and adds the likelihood only at sampling time. Practical TGD can use MPGD-, DPS-, or DAPS-style modules as approximate conditional reconstruction solvers. Independent best-of-N sampling improves robustness but spends computation uniformly across trajectories, even when some are poor early. Existing training-free samplers can be unreliable on challenging inverse problems.