Variational Diffusion Models (VDMs) are a new class of diffusion-based generative models that achieve state-of-the-art likelihoods on image density estimation benchmarks. Unlike other diffusion models, VDMs allow for efficient optimization of the noise schedule jointly with the rest of the model. The variational lower bound (VLB) simplifies to a remarkably short expression in terms of the signal-to-noise ratio of the diffused data, improving theoretical understanding of the model class. The continuous-time VLB is invariant to the noise schedule, except for the signal-to-noise ratio at its endpoints. This enables learning a noise schedule that minimizes the variance of the resulting VLB estimator, leading to faster optimization. Architectural improvements, including the use of Fourier features, lead to state-of-the-art likelihoods on image density estimation benchmarks, outperforming autoregressive models. VDMs can also be used in a bits-back compression scheme, achieving lossless compression rates close to the theoretical optimum. The code is available at https://github.com/google-research/vdm.Variational Diffusion Models (VDMs) are a new class of diffusion-based generative models that achieve state-of-the-art likelihoods on image density estimation benchmarks. Unlike other diffusion models, VDMs allow for efficient optimization of the noise schedule jointly with the rest of the model. The variational lower bound (VLB) simplifies to a remarkably short expression in terms of the signal-to-noise ratio of the diffused data, improving theoretical understanding of the model class. The continuous-time VLB is invariant to the noise schedule, except for the signal-to-noise ratio at its endpoints. This enables learning a noise schedule that minimizes the variance of the resulting VLB estimator, leading to faster optimization. Architectural improvements, including the use of Fourier features, lead to state-of-the-art likelihoods on image density estimation benchmarks, outperforming autoregressive models. VDMs can also be used in a bits-back compression scheme, achieving lossless compression rates close to the theoretical optimum. The code is available at https://github.com/google-research/vdm.