Direct Intermediate Initialization for Tilted Diffusion Samplers
The paper starts MCGDiff partway through the reverse process by sampling a softened clean-space posterior, then bridging those samples into the noisy tilted target.
That shortcut runs only the remaining SMC suffix, accepting a loss of asymptotic consistency for better finite-particle behavior. With MMPS as the initializer, the hybrid cuts sliced Wasserstein distance by about 2x on a structured Gaussian-mixture inverse problem at matched particle count. The gain is larger when the relevant posterior mode is rare under the prior, because resampling cannot recover a mode missing from the initial particles. On the standard benchmark, a control suggests much of the improvement comes from intermediate initialization itself, with conditioning adding more on harder cases. ArXiv · AI/CL/LG's note
That shortcut runs only the remaining SMC suffix, accepting a loss of asymptotic consistency for better finite-particle behavior. With MMPS as the initializer, the hybrid cuts sliced Wasserstein distance by about 2x on a structured Gaussian-mixture inverse problem at matched particle count. The gain is larger when the relevant posterior mode is rare under the prior, because resampling cannot recover a mode missing from the initial particles. On the standard benchmark, a control suggests much of the improvement comes from intermediate initialization itself, with conditioning adding more on harder cases. ArXiv · AI/CL/LG's note
score 4