We tested the efficacy of a training programme, delivered in virtual reality (VR), to modify the perceptual boundary between what participants classify as a fat versus a thin body. Three cohorts of 20 female volunteers with high body image concerns were recruited to two intervention groups and one control group. All participants completed a 4-day training programme in VR where they categorised a series of 3D models as either thin or fat; one intervention group was presented with the stimuli briefly, while the other group had no time limits imposed. Both intervention groups were given inflationary feedback to shift their categorisations of the stimulus models towards higher BMIs. Our results show that, compared to controls, both intervention groups shifted their categorical boundaries between Day 1 and follow-up on Day 14. Unlimited stimulus presentation times were associated with a larger training effect. Furthermore, both intervention groups experienced statistically significant reductions in their concerns about their own body shape, weight and eating habits. However, only in the group with longer stimulus presentation times were these reductions consistent with a clinically meaningful effect. These findings suggest that manipulating categorical perception in VR might provide a complementary addition to existing treatments for eating disorders.