Objective Non-alcoholic fatty liver disease (NAFLD) is potentially reversible. However, whether improvement of NAFLD leads to clinical benefits remains uncertain. We investigated the association between regression of NAFLD and the risk of incident diabetes in a longitudinal way. Methods A cohort of 11,260 adults who had NAFLD at in an initial exam, had the second evaluation for NAFLD status at 1~2 years from an initial exam were followed up for incident diabetes from 2001 and 2016. NAFLD was diagnosed with abdominal ultrasound. Results At baseline, NAFLD was regressed in 2,559 participants (22.7%). During 51,388 person-years of follow-up (median 4 years), 1,768 participants developed diabetes. The fully adjusted hazard ratio (HR) for incident diabetes in participants with regressed NAFLD compared to those with persistent NAFLD was 0.81 [95% confidence interval (CI) 0.72–0.92]. When assessed by NAFLD severity, among participants with a low NAFLD fibrosis score (NFS) (< -1.455), participants with regressed NAFLD had a lower risk of incident diabetes than those with persistent NAFLD (HR 0.77, 95% CI 0.68–0.88). However, in participants with an intermediate to high NFS (≥ -1.455), the risk of incident diabetes was not different between NAFLD regression and persistence groups (HR 1.12, 95% CI 0.82–1.51). Conclusions Regression of NAFLD was associated with decreased risk of incident diabetes compared to persistent NAFLD. However, the benefit was evident only for NAFLD patients with low NFS. This suggests that early intervention for NAFLD, before advanced fibrosis is present, may maximize the metabolic benefit from NAFLD regression.