Generative Vision Interview Questions, issue 24, Jul 3, 2026

The Trajectory Curvature Trap

Research Engineer interview at Google DeepMind, and the interviewer asks:

Your 1000-step diffusion teacher is too slow for production, so you distill it into a single-step student. The output is garbage. You swap in a bigger student. Still garbage. What’s actually broken?

Don’t say: The student doesn’t have enough capacity.

When one-step distillation fails, model size isn't your bottleneck, it's the hidden complexity of the ODE path, and how constant-difficulty sub-problems actually solve it.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.