The first idea is: you need a seismic model to take from the wells the bandwidth of the signal that you do not have on surface seismic acquisition.
Information required to build low-frequency models primarily comes from the well-log dataset and your seismic interpretation: your horizons.
On the other hand, you must build a geologic model first and compare the model to our seismic data with your real data: your horizons from your seismic track interpretation and your wells. You must compare then the results of this comparison between real and modeled data to iteratively update the model in such a way as to better match the seismic data. The low-frequency model is your geological background or framework to make geological sense of the seismic inversion output.
You know that seismic inversion - even acoustic, elastic, simultaneous, AVO, etc. - is the process of converting seismic reflection data into seismic impedance. And you know that impedance is the product of density and velocity. Therefore, you will go to the well-logs that have density and velocity to make the low-frequency model that the seismic does not have.
There are a lot of papers related to your question, I recommend reading:
Whitcombe D. and Hodgson L., 2007 Stablising the low frequencies, the Leading Edge 26, 66-72.
Kumar and Negi's, Hyderabad 2012: "Low frequency modeling and its impact on seismic inversion data".
The Inversion results may be dependent on the input model(s) for some inversion methods. Attention should be given to the test results when the thirds party commercial software is used.
@ Özdemir (2009), First Break for issues with the low frequency background modelling.
The answers before contain everything you need. I only want to repeat what is in the Mario's text: build a preliminary model from the information you have (geology, and overall logs), but keep your mind open, so that you refine it iteratively on the basis of the results of your inversion. And think also always about the assumptions you make when you perform the inversion, since it can explain eventual discrepancies.