Enhancing Decoupled Posterior Sampling with Data Consistency Guidance for Inverse Problems
In Proceedings of the 22nd International Conference on Intelligent Computing (ICIC 2026), Lecture Notes in Computer Science, pp. 167-179 (Oral Presentation)
Diffusion models solve inverse problems well through posterior sampling, but they accumulate error in the early steps. Decoupled posterior sampling methods address part of this, yet their reverse process ignores the measurement entirely, so those early errors carry forward and obstruct optimisation in later steps.
We propose Guided Decoupled Posterior Sampling (GDPS), which integrates a data consistency constraint into the reverse process. The constraint produces a smoother transition through the optimisation trajectory and converges more effectively toward the target distribution. The method extends to latent diffusion models and to Tweedie’s formula, which shows that it scales.
We evaluate GDPS on FFHQ and ImageNet across linear and nonlinear tasks, under both standard and challenging settings, where it reaches state-of-the-art performance.