Abstract Fracture networks act as critical pathways for groundwater flow and transport, yet their characterization remains challenging due to subsurface inaccessibility and stochastic complexity. Traditional inversion methods are computationally expensive and often fail to capture fracture heterogeneity accurately. We propose GenFrac, a pre‐trained generative AI method based on denoising diffusion, for autonomous inversion of fracture networks. This approach formulates inversion as a conditional denoising process using sparse, noisy observations and geological priors. A physics‐supervised screening step ensures generated fields are physically plausible. Applied to both theoretical geothermal and real unconfined aquifer cases, GenFrac improves reconstruction accuracy and provides a probabilistic framework for ensemble‐based uncertainty representation. The method enables efficient, generative parameter inversion conditioned on state observations, showing promise for broader application in subsurface flow and transport studies; formal uncertainty calibration and validation across diverse geological settings remain directions for future work.