Despite remarkable advances in protein structure prediction, a fundamental question remains unresolved: how do proteins fold from unfolded conformations into their native states? Here, we introduce PathDiffusion, a novel generative framework that simulates protein folding pathways using evolution-guided diffusion models. PathDiffusion first extracts structure-aware evolutionary information from 52 million predicted structures the AlphaFold database. Then an evolution-guided diffusion model with a dual-score fusion strategy is trained to generate high-fidelity folding pathways. Unlike existing deep learning methods, which primarily sample equilibrium ensembles, PathDiffusion explicitly models the temporal evolution of folding. On a benchmark of 52 proteins with experimentally validated folding pathways, PathDiffusion accurately reconstructs the order of folding events. We further demonstrate its versatility across four challenging applications: (1) recapitulating Anton’s molecular dynamics trajectory for 12 fast-folding proteins, (2) reproducing functionally important local folding-unfolding transitions in 20 proteins, (3) characterizing conformational ensembles of 50 intrinsically disordered proteins, and (4) resolving distinct folding mechanisms among 3 TIM-barrel proteins. We anticipate that PathDiffusion will be a valuable tool for probing protein folding mechanisms and dynamics at scale.