Generative Modeling Foundations
Likelihood, score-based modeling, DDPM forward/reverse process, denoising objectives, guidance and sampling loop trade-offs.
Что должен уметь кандидат
- Explain DDPM noising and reverse denoising at a conceptual and algorithmic level.
- Compare GAN/VAE/diffusion trade-offs by stability, diversity, quality and sampling cost.
- Reason about classifier-free guidance and why stronger guidance can create artifacts.
- Connect sample quality, number of denoising steps and latency.
Что спрашивают на собеседовании
- Why does predicting noise work in DDPM?
- What changes when using classifier-free guidance?
- Why are diffusion samplers slow?
Практическая задача
Implement or inspect a small DDPM/DDIM sampler and plot denoising trajectory and latency vs steps.
Source-grounded правило
Keep math claims tied to DDPM/LDM papers; avoid presenting one sampler family as universally superior.