There’s something I don’t understand. Improving the low-light performance of CMOS sensors is a long-standing, well-studied problem—one that academia and industry have poured millions of dollars into, from iPhone photography to fluorescence microscopy. How is it possible that a young team has “cracked it” with performance that appears to be 100 times better than the state of the art in their demo photos?
As I understand it, AI cannot surpass the fundamental limits of information theory. This means their method should achieve performance similar to “classical” denoising approaches—unless they are “filling in the information gaps” using AI, which sounds like a really bad idea, especially if they’re marketing their technology for military applications!
There’s something I don’t understand. Improving the low-light performance of CMOS sensors is a long-standing, well-studied problem—one that academia and industry have poured millions of dollars into, from iPhone photography to fluorescence microscopy. How is it possible that a young team has “cracked it” with performance that appears to be 100 times better than the state of the art in their demo photos?
As I understand it, AI cannot surpass the fundamental limits of information theory. This means their method should achieve performance similar to “classical” denoising approaches—unless they are “filling in the information gaps” using AI, which sounds like a really bad idea, especially if they’re marketing their technology for military applications!
website: https://deepnight.ai/
YC page: https://www.ycombinator.com/companies/deepnight
TC article: https://techcrunch.com/2025/02/27/yc-grad-deepnight-nabs-5-5...
Demos: https://www.youtube.com/watch?v=keq2ewcXpIc