Explain Denoising Diffusion models and forward/reverse processes.

WHAT THIS TESTS: The interviewer wants to know if you understand diffusion models as latent variable generative models that use a fixed noising process and a learned denoising process. They are checking whether you can distinguish diffusion from GANs, VAEs, and autoregressive models, and whether you grasp the role of iterative timesteps and noise prediction.

A GOOD ANSWER COVERS: First, the core idea: a diffusion model learns to reverse a gradual corruption process. It treats generation as navigating a random walk with drift through data space. Second, the forward process: you should describe it as a fixed Markov chain of T steps where Gaussian noise is added at each step according to a variance schedule, eventually turning any data point into pure isotropic Gaussian noise. Third, the reverse process: you should explain that a neural network, typically a U-Net, learns to predict the noise that was added at each step, enabling iterative denoising from pure noise back to a realistic sample. Fourth, training versus sampling: during training the model sees noisy data at random timesteps and learns to predict the noise; during sampling it starts from random noise and applies the learned denoising steps sequentially.

COMMON WRONG ANSWERS: A major red flag is describing diffusion as a single-step image-to-image network or an encoder-decoder like a VAE. Another mistake is saying the forward process is learned; it is fixed. Candidates also err by omitting the timestep conditioning, which tells the model how noisy the current input is, or by confusing the forward process with the reparameterization trick used in VAEs.

LIKELY FOLLOW-UPS: Expect the interviewer to ask how the model is trained, usually via a simple mean-squared-error loss between predicted and actual noise. They may ask why we sample from a Gaussian at each reverse step rather than taking a deterministic mean, which relates to the probabilistic nature of the reverse Markov chain. They might also ask about classifier-free guidance or the trade-off between sampling steps and quality.

ONE CONCRETE EXAMPLE: Imagine a 256 by 256 image of a cat. In the forward process, over one thousand steps, tiny amounts of Gaussian noise are added until the image becomes indistinguishable from static. In the reverse process, the model receives pure static and a timestep t equals one thousand, predicts the noise component, subtracts it, and repeats for nine hundred ninety-nine more steps until a sharp cat image emerges.