Policies and interventions that help individuals defend against and cope with the most damaging natural disaster in the US require flood risk estimates over large spatial scales. These rely on Federal Emergency Management Agency (FEMA) and United States Army Corps of Engineers (USACE) synthetic, expert based, depth damage functions (DDFs) that lack empirical validation even though this transfer function is often found as the most important component of the flood loss modeling process. Here, we perform a validation of damage functions on a dataset of 40 times more unique events and 10,000 times more observations than the next most expansive empirical effort. In our analysis, we develop a range of empirical functions and test them against FEMA/USACE DDF prediction benchmarks in external validation. Finally, we apply the benchmark and empirical models to a dataset of over 140 million single family homes in the US, with flood loss drivers specified at the property level, to observe the implications of estimator differences from property to aggregate scales. Internal to the loss data, we find that each estimator yields approximately mean unbiased estimates. However, they disagree substantially at the property level since losses are highly heterogeneous, even conditioned on the set of observables available for use in large scale flood loss estimation. These differences have surprising implications on the estimated flood loss distribution to single family homes in the US. Flood loss estimates aggregated from census block to county and state levels agree to a greater extent than expected across estimators in terms of absolute and rank ordering of losses. Property level losses agree to a lesser extent, but more than expected based off validation internal to the loss data. The level of agreement at any scale varies over space (i.e. FEMA Region). While losses are sensitive to the choice of damage estimator, our results provide strong evidence that flood risk management decisions at aggregate scales may be reliably informed by flood loss estimates. We make our preferred and easy to use empirical damage functions openly available and discuss how they could be used in large scale flood loss estimation to improve the understanding of flood risk to single family homes in the US and opportunities to effectively manage that risk. Future work will focus on improving the validity of damage functions examined here, generating predictions under uncertainty, and understanding the scales at which predictions can be applied reliably in different settings.