A lot of things which look good in the lab don't make it to production because issues not highlighted in the original research though.

I've seen this in software as well, trying to implement novel algorithms from papers.

For example, one raytracing denoising algorithm I tried worked wonders, except it required the entire rendering to be serialized (raytracing is "embarrassingly parallel"), alternatively use a per-pixel instance of the data structure, leading to many GB of extra memory required for larger images. Results looked great in the paper, not so much in the real world...