Eugene Yan's brief comparison walks through two families of generative models: autoencoders (variational and denoising) and diffusers. Autoencoders are trained to compress an input into a lower-dimensional latent space and then reconstruct it, with variational autoencoders adding a probabilistic twist to the latent distribution. Denoising autoencoders, meanwhile, learn to remove noise from corrupted inputs, which is a related but distinct objective.
Diffusion models take a different route: they start with clean data and progressively add noise over many steps, then learn to reverse that process to generate new samples from pure noise. This means diffusers are inherently iterative, while autoencoders typically map directly from input to output in a single pass.
The source notes that while both can be used for generation, they have different strengths. Autoencoders are often faster and more compact, but diffusers tend to produce higher-fidelity samples, especially for images. The comparison is brief, so it doesn't dive into training details or recent hybrids, but it offers a clear starting point for understanding the trade-offs.