Point sources such as power plants and oil and gas facilities contribute over 40% of global anthropogenic CO₂ emissions, yet reliably monitoring emissions from these sources remains a challenge. Inverse modeling of atmospheric CO2 measurements has become a widely-used method for quantifying these sources. Commonly used inversion methods such as Gaussian Plume Modeling or Cross-Sectional Flux estimation do not explicitly simulate the CO2 emission plumes and rely on steady-state assumptions. Here, we use the Weather Research and Forecasting (WRF) model to simulate the CO2 plume from point sources to quantify CO2 emissions across a range of scales within a Bayesian inversion framework. At the kilometer scale, we use satellite observations from the Orbiting Carbon Observatory-3 (OCO-3) to quantify emissions from the Bełchatów power plant. At the sub-kilometer scale, we use the WRF large eddy simulation (WRF-LES) with observations from an uncrewed aerial vehicle (UAV) to estimate emissions from an oil and gas facility. We find that the inversion can reliably quantify power plant emissions at the km-scale. However, at the sub-km scale, although the inversion can recover accurate emissions, capturing point-to-point consistency between the model and observations is challenging. We also explore the use of a machine learning (ML) emulator (U-Net-Shuffle) to model the CO2 plume. The ML emulator directly maps low-resolution (1.6 km) meteorology to super-resolution (100 m) CO₂ footprints (100 m), greatly reducing the computational cost of the high-resolution modeling. An observing system simulation experiment (OSSE) inversion test is conducted and recovers the point source emissions using the ML predicted footprints. Our results demonstrate that high-resolution modeling of CO2 plumes from point sources in the context of a Bayesian inversion framework offers the potential for quantifying CO2 emissions from these sources under a broader range of meteorological conditions than is possible with widely-used methods.