DDIM: faster diffusion sampling

WHAT THIS TESTS The interviewer wants to know whether you understand why DDPM is slow and how DDIM speeds it up while reusing the same trained model.

A GOOD ANSWER COVERS DDPM learns to reverse a Markovian forward diffusion one small step at a time, so generating a sample requires many sequential denoising steps, often around a thousand, which is expensive. DDIM observes that the training objective only constrains the marginal noise distributions, not the specific Markov chain, so it constructs a family of non-Markovian generative processes that share the same noise-prediction network. By making the reverse process deterministic, with zero added noise, DDIM can jump across multiple timesteps at once, sampling in perhaps twenty to fifty steps. Crucially, no retraining is needed: the same DDPM-trained network is reused with a different sampling rule. Determinism also yields reproducible outputs from a fixed latent and smooth latent-space interpolation.

COMMON WRONG ANSWERS Saying DDIM needs its own training run or a different loss; it reuses the DDPM network. Another error is claiming DDIM is free with no downside; aggressively few steps lose fine detail and can soften or distort images, so there is a real speed-quality trade-off.

LIKELY FOLLOW-UPS Expect questions on the stochasticity parameter eta that interpolates between DDIM and DDPM, on why determinism enables interpolation, on higher-order solvers like DPM-Solver, and on distillation methods for one-step generation.

ONE CONCRETE EXAMPLE Using the same trained model, DDPM might need a thousand steps to render a crisp portrait, while DDIM produces a comparable image in fifty deterministic steps, roughly twenty times faster. Push it to five steps and the face still forms but loses fine texture and sharpness, illustrating the trade-off.