{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,5,2]],"date-time":"2026-05-02T22:32:11Z","timestamp":1777761131072,"version":"3.51.4"},"reference-count":73,"publisher":"Emerald","issue":"2","content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"published-print":{"date-parts":[[2025,4,28]]},"abstract":"<jats:p>Despite the growing interest in diffusion models, gaining a deep understanding of the model class remains an elusive endeavour, particularly for the uninitiated in non-equilibrium statistical physics. Thanks to the rapid rate of progress in the field, most existing work on diffusion models focuses on either applications or theoretical contributions. Unfortunately, the theoretical material is often inaccessible to practitioners and new researchers, leading to a risk of superficial understanding in ongoing research. Given that diffusion models are now an indispensable tool, a clear and consolidating perspective on the model class is needed to properly contextualize recent advances in generative modelling and lower the barrier to entry for new researchers. To that end, we revisit predecessors to diffusion models, such as hierarchical latent variable models, and synthesize a holistic perspective using only directed graphical modelling and variational inference principles. The resulting narrative is easier to follow as it imposes fewer prerequisites on the average reader relative to the view from non-equilibrium thermodynamics or stochastic differential equations.<\/jats:p>","DOI":"10.1561\/0600000113","type":"journal-article","created":{"date-parts":[[2025,4,28]],"date-time":"2025-04-28T03:46:19Z","timestamp":1745811979000},"page":"76-170","source":"Crossref","is-referenced-by-count":2,"title":["Demystifying Variational Diffusion Models"],"prefix":"10.1108","volume":"17","author":[{"given":"Fabio","family":"De Sousa Ribeiro","sequence":"first","affiliation":[{"name":"Imperial College London"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Ben","family":"Glocker","sequence":"additional","affiliation":[{"name":"Imperial College London"}],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"140","published-online":{"date-parts":[[2025,4,28]]},"reference":[{"key":"2026032615065987100_ref001","article-title":"Building normalizing flows with stochastic interpolants","volume-title":"The Eleventh International Conference on Learning Representations","author":"Albergo","year":"2023"},{"issue":"3","key":"2026032615065987100_ref002","doi-asserted-by":"crossref","first-page":"313","DOI":"10.1016\/0304-4149(82)90051-5","article-title":"Reverse-time diffusion equation models","volume":"12","author":"Anderson","year":"1982","journal-title":"Stochastic Processes and their Applications"},{"key":"2026032615065987100_ref003","volume-title":"Deep Learning: Foundations and Concepts","author":"Bishop","year":"2023"},{"issue":"518","key":"2026032615065987100_ref004","doi-asserted-by":"crossref","first-page":"859","DOI":"10.1080\/01621459.2017.1285773","article-title":"Variational inference: A review for statisticians","volume":"112","author":"Blei","year":"2017","journal-title":"Journal of the American statistical Association"},{"key":"2026032615065987100_ref005","article-title":"Importance weighted autoencoders","volume-title":"arXiv preprint arXiv:1509.00519","author":"Burda","year":"2015"},{"key":"2026032615065987100_ref006","article-title":"Very deep vaes generalize autoregressive models and can outperform them on images","volume-title":"International Conference on Learning Representations","author":"Child","year":"2020"},{"key":"2026032615065987100_ref007","first-page":"11472","article-title":"Perception prioritized training of diffusion models","volume-title":"Proceedings of the IEEE\/CVF Conference on Computer Vision and Pattern Recognition","author":"Choi","year":"2022"},{"key":"2026032615065987100_ref008","first-page":"7390","volume-title":"Proceedings of the 40th International Conference on Machine Learning","author":"De Sousa Ribeiro","year":"2023"},{"key":"2026032615065987100_ref009","first-page":"8780","article-title":"Diffusion models beat gans on image synthesis","volume":"34","author":"Dhariwal","year":"2021","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref010","article-title":"Continuous diffusion for categorical data","author":"Dieleman","year":"2022","journal-title":"arXiv preprint arXiv:2211.15089"},{"key":"2026032615065987100_ref011","first-page":"8489","volume-title":"International conference on machine learning","author":"Du","year":"2023"},{"key":"2026032615065987100_ref012","article-title":"Efficient diffusion training via min-snr weighting strategy","author":"Hang","year":"2023","journal-title":"arXiv preprint arXiv:2303.09556"},{"key":"2026032615065987100_ref013","article-title":"Gans trained by a two time-scale update rule converge to a local nash equilibrium","volume":"30","author":"Heusel","year":"2017","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref014","first-page":"6840","article-title":"Denoising diffusion probabilistic models","volume":"33","author":"Ho","year":"2020","journal-title":"Advances in neural information processing systems"},{"issue":"1","key":"2026032615065987100_ref015","first-page":"2249","article-title":"Cascaded diffusion models for high fidelity image generation","volume":"23","author":"Ho","year":"2022","journal-title":"The Journal of Machine Learning Research"},{"key":"2026032615065987100_ref016","article-title":"Stochastic variational inference","volume-title":"Journal of Machine Learning Research","author":"Hoffman","year":"2013"},{"key":"2026032615065987100_ref017","article-title":"Elbo surgery: Yet another way to carve up the variational evidence lower bound","volume":"1","author":"Hoffman","year":"2016","journal-title":"Workshop in Advances in Approximate Bayesian Inference, NIPS"},{"key":"2026032615065987100_ref018","first-page":"13 213","volume-title":"Proceedings of the 40th International Conference on Machine Learning","author":"Hoogeboom","year":"2023"},{"key":"2026032615065987100_ref019","article-title":"Simpler diffusion (sid2): 1.5 fid on imagenet512 with pixel-space diffusion","volume-title":"arXiv preprint arXiv:2410.19324","author":"Hoogeboom","year":"2024"},{"key":"2026032615065987100_ref020","first-page":"8867","volume-title":"International conference on machine learning","author":"Hoogeboom","year":"2022"},{"key":"2026032615065987100_ref021","first-page":"22 863","article-title":"A variational perspective on diffusion-based generative models and score matching","volume":"34","author":"Huang","year":"2021","journal-title":"Advances in Neural Information Processing Systems"},{"issue":"4","key":"2026032615065987100_ref022","article-title":"Estimation of non-normalized statistical models by score matching","volume":"6","author":"Hyv\u00e4rinen","year":"2005","journal-title":"Journal of Machine Learning Research"},{"issue":"1","key":"2026032615065987100_ref023","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1007\/s10463-023-00884-4","article-title":"Identifiability of latent-variable and structural-equation models: From linear to nonlinear","volume":"76","author":"Hyv\u00e4rinen","year":"2024","journal-title":"Annals of the Institute of Statistical Mathematics"},{"key":"2026032615065987100_ref024","doi-asserted-by":"crossref","first-page":"183","DOI":"10.1023\/A:1007665907178","article-title":"An introduction to variational methods for graphical models","volume":"37","author":"Jordan","year":"1999","journal-title":"Machine learning"},{"key":"2026032615065987100_ref025","first-page":"5006","volume-title":"International Conference on Machine Learning","author":"Jun","year":"2020"},{"key":"2026032615065987100_ref026","first-page":"26 565","article-title":"Elucidating the design space of diffusion-based generative models","volume":"35","author":"Karras","year":"2022","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026032615065987100_ref027","first-page":"24 174","article-title":"Analyzing and improving the training dynamics of diffusion models","author":"Karras","year":"2024","journal-title":"Proceedings of the IEEE\/CVF Conference on Computer Vision and Pattern Recognition"},{"key":"2026032615065987100_ref028","first-page":"21 696","article-title":"Variational diffusion models","volume":"34","author":"Kingma","year":"2021","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref029","article-title":"Understanding the diffusion objective as a weighted integral of elbos","volume-title":"arXiv preprint arXiv:2303.00848","author":"Kingma","year":"2023"},{"key":"2026032615065987100_ref030","article-title":"Auto-encoding variational bayes","volume-title":"arXiv preprint arXiv:1312.6114","author":"Kingma","year":"2013"},{"issue":"4","key":"2026032615065987100_ref031","doi-asserted-by":"crossref","first-page":"307","DOI":"10.1561\/2200000056","article-title":"An introduction to variational autoencoders","volume":"12","author":"Kingma","year":"2019","journal-title":"Foundations and Trends\u00ae in Machine Learning"},{"key":"2026032615065987100_ref032","article-title":"Improved variational inference with inverse autoregressive flow","volume":"29","author":"Kingma","year":"2016","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref033","volume-title":"Tutorial on denoising diffusionbased generative modeling: Foundations and applications","author":"Kreis","year":"2022"},{"key":"2026032615065987100_ref034","article-title":"Flow matching for generative modeling","volume-title":"The Eleventh International Conference on Learning Representations","author":"Lipman","year":"2023"},{"key":"2026032615065987100_ref035","article-title":"Flow straight and fast: Learning to generate and transfer data with rectified flow","volume-title":"The Eleventh International Conference on Learning Representations","author":"Liu","year":"2023"},{"key":"2026032615065987100_ref036","article-title":"Understanding diffusion models: A unified perspective","volume-title":"arXiv preprint arXiv:2208.11970","author":"Luo","year":"2022"},{"key":"2026032615065987100_ref037","article-title":"Biva: A very deep hierarchy of latent variables for generative modeling","volume":"32","author":"Maal\u00f8e","year":"2019","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref038","first-page":"1445","volume-title":"International conference on machine learning","author":"Maal\u00f8e","year":"2016"},{"key":"2026032615065987100_ref039","article-title":"Adversarial autoencoders","volume-title":"arXiv preprint arXiv:1511.05644","author":"Makhzani","year":"2015"},{"key":"2026032615065987100_ref040","article-title":"Improved autoregressive modeling with distribution smoothing","volume-title":"International Conference on Learning Representations","author":"Meng","year":"2020"},{"key":"2026032615065987100_ref041","article-title":"Measuring axiomatic soundness of counterfactual image models","volume-title":"The Eleventh International Conference on Learning Representations","author":"Monteiro","year":"2022"},{"key":"2026032615065987100_ref042","article-title":"Continuous diffusion for mixed-type tabular data","volume-title":"arXiv preprint arXiv:2312.10431","author":"Mueller","year":"2023"},{"key":"2026032615065987100_ref043","first-page":"8162","volume-title":"International Conference on Machine Learning","author":"Nichol","year":"2021"},{"key":"2026032615065987100_ref044","first-page":"16 784","volume-title":"International Conference on Machine Learning","author":"Nichol","year":"2022"},{"key":"2026032615065987100_ref045","first-page":"324","volume-title":"International conference on machine learning","author":"Ranganath","year":"2016"},{"key":"2026032615065987100_ref046","article-title":"Semi-supervised learning with ladder networks","volume":"28","author":"Rasmus","year":"2015","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref047","unstructured":"D. J.\n              Rezende\n            \n          , Short notes on divergence measures, 2018. url: https:\/\/danilorezende.com\/wp-content\/uploads\/2018\/07\/divergences.pdf."},{"key":"2026032615065987100_ref048","first-page":"1278","volume-title":"International conference on machine learning","author":"Rezende","year":"2014"},{"key":"2026032615065987100_ref049","article-title":"Taming vaes","volume-title":"arXiv preprint arXiv:1810.00597","author":"Rezende","year":"2018"},{"key":"2026032615065987100_ref050","article-title":"Generative modelling with inverse heat dissipation","volume-title":"The Eleventh International Conference on Learning Representations","author":"Rissanen","year":"2023"},{"key":"2026032615065987100_ref051","first-page":"10 684","article-title":"High-resolution image synthesis with latent diffusion models","volume-title":"Proceedings of the IEEE\/CVF conference on computer vision and pattern recognition","author":"Rombach","year":"2022"},{"key":"2026032615065987100_ref052","article-title":"Align your steps: Optimizing sampling schedules in diffusion models","volume-title":"arXiv preprint arXiv:2404.14507","author":"Sabour","year":"2024"},{"key":"2026032615065987100_ref053","first-page":"36 479","article-title":"Photorealistic text-to-image diffusion models with deep language understanding","volume":"35","author":"Saharia","year":"2022","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026032615065987100_ref054","article-title":"Improved techniques for training gans","volume":"29","author":"Salimans","year":"2016","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref055","article-title":"Should EBMs model the energy or the score?","volume-title":"Energy Based Models Workshop - ICLR 2021","author":"Salimans","year":"2021"},{"key":"2026032615065987100_ref056","article-title":"Progressive distillation for fast sampling of diffusion models","volume-title":"International Conference on Learning Representations","author":"Salimans","year":"2022"},{"key":"2026032615065987100_ref057","first-page":"1218","volume-title":"International conference on machine learning","author":"Salimans","year":"2015"},{"key":"2026032615065987100_ref058","article-title":"Using ornstein-uhlenbeck process to understand denoising diffusion probabilistic model and its noise schedules","volume-title":"arXiv preprint arXiv:2311.17673","author":"Santos","year":"2023"},{"key":"2026032615065987100_ref059","first-page":"20 141","volume-title":"International Conference on Machine Learning","author":"Shu","year":"2022"},{"key":"2026032615065987100_ref060","first-page":"2256","volume-title":"International conference on machine learning","author":"Sohl-Dickstein","year":"2015"},{"key":"2026032615065987100_ref061","article-title":"Ladder variational autoencoders","volume":"29","author":"S\u00f8nderby","year":"2016","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref062","first-page":"1415","article-title":"Maximum likelihood training of score-based diffusion models","volume":"34","author":"Song","year":"2021","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026032615065987100_ref063","article-title":"Generative modeling by estimating gradients of the data distribution","volume":"32","author":"Song","year":"2019","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref064","first-page":"12 438","article-title":"Improved techniques for training scorebased generative models","volume":"33","author":"Song","year":"2020","journal-title":"Advances in neural information processing systems"},{"key":"2026032615065987100_ref065","article-title":"How to train your energy-based models","volume-title":"arXiv preprint arXiv:2101.03288","author":"Song","year":"2021"},{"key":"2026032615065987100_ref066","article-title":"Score-based generative modeling through stochastic differential equations","volume-title":"International Conference on Learning Representations","author":"Song","year":"2021"},{"key":"2026032615065987100_ref067","first-page":"1214","volume-title":"International Conference on Artificial Intelligence and Statistics","author":"Tomczak","year":"2018"},{"key":"2026032615065987100_ref068","first-page":"19 667","article-title":"Nvae: A deep hierarchical variational autoencoder","volume":"33","author":"Vahdat","year":"2020","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026032615065987100_ref069","first-page":"11 287","article-title":"Score-based generative modeling in latent space","volume":"34","author":"Vahdat","year":"2021","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026032615065987100_ref070","doi-asserted-by":"crossref","first-page":"143","DOI":"10.1016\/B978-0-12-802806-3.00008-7","volume-title":"Advances in independent component analysis and learning machines","author":"Valpola","year":"2015"},{"issue":"7","key":"2026032615065987100_ref071","doi-asserted-by":"crossref","first-page":"1661","DOI":"10.1162\/NECO_a_00142","article-title":"A connection between score matching and denoising autoencoders","volume":"23","author":"Vincent","year":"2011","journal-title":"Neural computation"},{"key":"2026032615065987100_ref072","doi-asserted-by":"crossref","first-page":"1096","DOI":"10.1145\/1390156.1390294","article-title":"Extracting and composing robust features with denoising autoencoders","volume-title":"Proceedings of the 25th international conference on Machine learning","author":"Vincent","year":"2008"},{"key":"2026032615065987100_ref073","doi-asserted-by":"crossref","DOI":"10.1007\/978-0-387-21736-9","volume-title":"All of statistics: a concise course in statistical inference","author":"Wasserman","year":"2004"}],"container-title":["Foundations and Trends\u00ae in Computer Graphics and Vision"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/www.emerald.com\/ftcgv\/article-pdf\/17\/2\/76\/11044772\/0600000113en.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"https:\/\/www.emerald.com\/ftcgv\/article-pdf\/17\/2\/76\/11044772\/0600000113en.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,4,29]],"date-time":"2026-04-29T14:09:13Z","timestamp":1777471753000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.emerald.com\/ftcgv\/article\/17\/2\/76\/1328356\/Demystifying-Variational-Diffusion-Models"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2025,4,28]]},"references-count":73,"journal-issue":{"issue":"2","published-print":{"date-parts":[[2025,4,28]]}},"URL":"https:\/\/doi.org\/10.1561\/0600000113","relation":{},"ISSN":["1572-2740","1572-2759"],"issn-type":[{"value":"1572-2740","type":"print"},{"value":"1572-2759","type":"electronic"}],"subject":[],"published":{"date-parts":[[2025,4,28]]}}}