While writing a raytracing engine for my university project (a fairly long and complex program in C++), there was a subtle bug that, under very specific conditions, the ray energy calculation would return 0 or NaN, and the corresponding pixel would be slightly dimmer than it should be.

Now you might think that this is a trifling problem, but when it happened to random pixels across the screen at random times it would manifest as noise, and as you might know, people who render stuff Absolutely. Hate. Noise. It wouldn't do. Not acceptable.

So I worked at that thing for three whole days and finally located the bug, a tiny gotcha-type thing in a numerical routine in one corner of the module that handled multiple importance sampling (basically, mixing different sampling strategies).

Frustrating, exhausting, and easily the most gruelling bug hunt I've ever done. Utterly worth it when I fixed it. And what's even better, I found and squashed two other bugs I hadn't even noticed, lol