Consecutive Posterior Fusion for Diffusive Recovery of Unobservable Image Structures
CPF-DDNM uses consecutive posterior estimates to sharpen diffusion recovery where the measurements cannot see.
The paper presents an inference-time method for ill-posed imaging tasks, built inside DDNM, that does not require retraining or extra denoiser calls. Its fusion step is described as preserving the measurement-determined component while changing only the prior-driven null-space estimate. The authors give a geometric interpretation and local error analysis for the time-dependent fusion coefficient, including extrapolation. Experiments cover sparse-view and simulated low-dose CT, plus medical image super-resolution, where the method improves over DDNM and remains competitive with other diffusion inverse solvers. ArXiv · AI/CL/LG's note
The paper presents an inference-time method for ill-posed imaging tasks, built inside DDNM, that does not require retraining or extra denoiser calls. Its fusion step is described as preserving the measurement-determined component while changing only the prior-driven null-space estimate. The authors give a geometric interpretation and local error analysis for the time-dependent fusion coefficient, including extrapolation. Experiments cover sparse-view and simulated low-dose CT, plus medical image super-resolution, where the method improves over DDNM and remains competitive with other diffusion inverse solvers. ArXiv · AI/CL/LG's note
score 4