Diffusion, Flow Matching and DiT
From U-Net latent diffusion to Diffusion Transformers, flow matching, rectified flow and modern high-resolution generative model design.
Что должен уметь кандидат
- Compare pixel-space and latent-space generation by compute and representation.
- Explain DiT as transformer over latent patches and why scale matters.
- Describe flow matching as vector-field regression at a high level.
- Understand why few-step/flow-style methods matter for production latency.
Что спрашивают на собеседовании
- Why did DiT replace U-Net in some modern systems?
- What is the difference between diffusion sampling and ODE/flow generation?
- Why does latent-space training reduce compute?
Практическая задача
Run a latent diffusion pipeline, vary scheduler settings and write a short report on latency/quality trade-offs.
Source-grounded правило
Sora/SD3-style reports are useful but not always fully reproducible; label them as technical reports or product reports where appropriate.