Generative models often struggle when the target distribution has limited data. The authors of this paper ask whether related distributions can help: they develop a diffusion transfer framework that learns a shared score from pooled data, then adds a target-specific offset. This two-stage decomposition is designed to exploit common structure while still adapting to the target's unique features.

The theory, built under Besov regularity and a stable hierarchical composition model for the offset, yields convergence rates that separate the cost of learning the shared component from the cost of learning the offset. The key result is that when the offset is simple, the risk bound improves over fitting the target separately. The authors also prove that the offset remains bounded and jointly regular as diffusion time approaches zero, which lets the lower time cutoff vanish polynomially and gives total-variation guarantees for the resulting density estimator.

Synthetic simulations and image benchmarks provide a proof of concept, though the paper does not compare against a wide range of baselines. Since this is the only source, there are no conflicting accounts to note; the claims stand as presented in the preprint.