Learning 3D human pose prior is essential to human centered AI. Here, we present GFPose, a versatile frame work to model plausible 3D human poses for various appli cations. At the core of GFPose is a time-dependent score network, which estimates the gradient on each body joint and progressively denoises the perturbed 3D human pose to match a given task specification. During the denois ing process, GFPose implicitly incorporates pose priors in gradients and unifies various discriminative and genera tive tasks in an elegant framework. Despite the simplic ity, GFPose demonstrates great potential in several down stream tasks. Our experiments empirically show that 1) as a multi-hypothesis pose estimator, GFPose outperforms exist ing SOTAs by 20% on Human3.6M dataset. 2) as a single hypothesis pose estimator, GFPose achieves comparable re sults to deterministic SOTAs, even with a vanilla backbone. 3) GFPose is able to produce diverse and realistic samples in pose denoising, completion and generation tasks.