Too Faithful and Not Faithful Enough: Generative Models from VAEs to GANs and Flows

Deep End Era. Engineering with Sankalp · Oct 28, 2026 · 55m

A fictional insurer's synthetic claims data matches every histogram, yet one record in two hundred is a near-copy of a real customer and the rare fraud patterns have vanished. Statistician Tomasz Wren shows data scientist Ines Okafor why that is the central tension of generative modelling. Part seventeen of The Deep End's Machine Learning series covers autoregressive models, VAEs, GANs and normalising flows: what each optimises, how each fails, how to grade a generator, and when not to use one. Diffusion (ep96) is linked, not repeated. Cross-references: ep15, ep96, ep101, ep160, parts 1, 2…