Diffusion-based inverse problem solvers fail not because of complex measurement models, but because they underestimate or overestimate posterior uncertainty at intermediate timesteps—a problem you can now diagnose and fix.
This paper explains why diffusion models sometimes fail at solving inverse problems (like image reconstruction). The authors show that popular methods use approximations that incorrectly estimate how spread-out the solution should be at intermediate steps, leading to hallucinations or wrong answers. They provide a diagnostic tool to test when and why these failures happen.