Rényi Diffusion Models

The choice of training objective is central to diffusion-based generative modeling in terms of both sample quality and distribution coverage. While standard maximum likelihood training provides a principled objective with strong theoretical grounding, empirical studies indicate that previous training objectives in diffusion models often face an inverse correlation between likelihood optimization and perceptual evaluations. We propose the Rényi diffusion model, a unified generative framework that formulates training objectives using Rényi divergence. This yields a generalized score matching objective providing explicit control over the trade-off between sample quality and distribution coverage. Experiments demonstrate improved balance between density estimation and sample generation performances across multiple datasets without modifying model architectures or sampling procedures.

This work studies how to train diffusion models so that they both generate visually good samples and cover the full data distribution well. Standard likelihood-based training is theoretically well motivated, but in practice it can conflict with perceptual sample quality: models that achieve better likelihood may not always produce better-looking samples. To address this issue, we introduce the Rényi diffusion model, a unified framework that derives diffusion training objectives from Rényi divergence. This leads to a generalized score matching objective with a tunable parameter that controls the trade-off between sample quality and distribution coverage. Our experiments show that this approach improves the balance between density estimation and sample generation across multiple datasets, while keeping the same model architectures and sampling procedures.