Abstract
We put forward the concept of work extraction from thermal noise by phase-sensitive (homodyne) measurements of the noisy input followed by (outcome-dependent) unitary manipulations of the post-measured state. For optimized measurements, noise input with more than one quantum on average is shown to yield heat-to-work conversion with efficiency and power that grow with the mean number of input quanta, detector efficiency and its inverse temperature. This protocol is shown to be advantageous compared to common models of information and heat engines.