papersTODAY 04:00 UTC
Backward SDE Approach Targets Physics Consistency in Diffusion Models
A new arXiv preprint proposes using backward stochastic differential equations to enforce physics and measurement consistency when score-based diffusion models are applied to inverse problems. The authors argue that existing methods, such as heuristic guidance, periodic projections, or task-specific conditional training, are less principled. The work positions the approach as a general way to keep pretrained diffusion priors aligned with physical constraints.