NOBUGS 2026

Ptychography is a lensless, high-resolution advanced imaging technique widely used at modern synchrotron facilities. The integration of deep learning methods has partially addressed the issues of high computational cost and strict data overlap requirements associated with traditional algorithms. However, existing discriminative learning models still face limitations in reconstruction accuracy and training stability. In this talk, we present PtychoDiffusion, a physics-guided diffusion posterior sampling framework for high-fidelity ptychographic reconstruction. The framework leverages the strengths of diffusion models by learning the denoise process from random noise to target sample. After that, we innovatively incorporate physical principles into the model by embedding classical iterative projection algorithms into the denoising steps, achieving a combination of generative priors and physical constraints, thereby significantly enhancing reconstruction reliability. Numerical experiments using simulated ptychographic datasets demonstrate that PtychoDiffusion delivers higher reconstruction accuracy and more consistent image details under various conditions, with particularly notable advantages in recovering detail structures and handling low-overlap data. Moreover, preliminary validation on experimental ptychography data further suggests the potential of the framework for practical beamline reconstruction workflows. These results indicate that physics-guided diffusion framework offers a principled and effective paradigm for unsupervised ptychography reconstruction and shows promising application potential.