Product preview

Computer vision interviews used to stop at detection and a little geometry. That changed once video generation became a product line. A single loop for a vision or generative media role can now cover backprop by hand, NMS in Python, the flow matching loss, and a system design round on serving text-to-video at scale.


We pulled together the 55 questions that show up most often across candidate reports, prep forums and the skills listed in current job postings. Thirty of them come with the answer we would expect from a strong senior candidate: the core idea first, then the trade-off or the number that shows real experience. Try answering the other twenty-five yourself before checking. Read each answer, then try saying it out loud in under two minutes. That is roughly the time you get before the follow-up.


How these loops are usually shaped


Most loops for vision and video roles have five parts. There is a coding round, often ML coding in NumPy or PyTorch rather than pure algorithms. There is a depth round on vision or generative models. There is an ML system design round. There is a project deep dive where you defend your own work. Finally there is a behavioural round. Camera and on-device teams often add C++ and performance questions. Research-leaning teams spend more time on diffusion internals and paper discussion.


The questions below follow roughly the order in which a loop builds up: fundamentals, classical vision, perception, architectures, video, generation, then production.


Meet the creator

Computer Vision & Video Gen AI The Big Tech Interview Insider

$19
$100(81% off)