Diffusion models trade sampling speed for stable training and broad mode coverage relative to GANs and VAEs

GAN training is unstable due to the adversarial minimax game, and GANs tend to underrepresent rare modes of the data distribution. VAEs optimise a surrogate (ELBO) rather than the true likelihood, limiting sample quality. Flow-based models require architectures with invertible transforms at every layer, severely constraining the model family. Diffusion models avoid all three: training is stable (the forward process is fixed, not learned), the objective is a well-defined variational bound on log-likelihood, and the model architecture is unconstrained. The cost is slow sequential sampling — generating one sample requires hundreds to thousands of network evaluations. New techniques (DDIM, consistency models) have dramatically reduced this gap, but diffusion sampling is still slower than a single GAN forward pass.

Examples

Sampling 50k images of size 32x32 from DDPM takes about 20 hours on a 2080 Ti; a GAN does it in under a minute. DDIM with 50 steps reduces DDPM sampling time by roughly 20x at comparable quality.

Assessment

List one strength and one weakness of diffusion models relative to GANs. Then explain which practical technique reduces the core weakness.