Score-Control for Hallucination Reduction in Diffusion Models

Score Control for Hallucination Reduction in Diffusion Models, from the University at Buffalo, argues that hallucinations in diffusion generators arise when the learned score field is overly smooth, which leaks nonzero probability mass into off-manifold regions. The paper formalizes this with a lower bound on model density outside data support that depends on the score Lipschitz constant L and score magnitude bound S, then introduces Variance-Guided Score Modulation, or VSM, an architecture-agnostic training objective that penalizes small score Jacobians using the variance-learning head from I-DDPM. A time-dependent schedule amplifies this penalty near the final denoising steps, where hallucinations are most likely. On synthetic and real image benchmarks, including 1D and 2D Gaussian mixtures, Hands-11K, MNIST, Shapes, Cards, ChessImages, and ImageNet-1K, VSM consistently reduces hallucination rates while preserving or improving fidelity and diversity. On the synthetic mixtures, it lowers score RMSE and hallucination rate; on Hands-11K, it cuts hallucinations from 23.33% to 5.15% for DDPM and from 29.50% to 21.15% for text-conditioned LDM; on Cards, it reduces hallucinations from 22.41% to 2.33%. The paper also introduces ChessImages, with an extreme semantic space of about 10^44 valid board states, to enable rule-checkable hallucination analysis and show that VSM increases valid novel generations rather than merely memorizing training boards.