I am generating an unbinned data set from a model which them I am fitting it to it. When I am plotting the model on the data in a RooPlot this is working fine. Afterwards I need to plot my data with variable bins so I create a RooDataHist from a TH1*. Then I plot on the top of that the RooAbsPdf I used in the unbinned PDF in order to do a chi2 calculation. However, the plotting of the PDF on the top of the RooDataHist with the variable bins seems to get confused with the normalization.

I guess you suffer from the problem that PDFs are densities, but histograms display counts, i.e. your histograms needs to be normalised using the bin width.
I remember that somebody asked for this, so I proposed some code in this post

to normalise the data by bin width. Could you try if this gives the expected result? If yes, I can make this option more easily available when plotting a histogram.
The plotting code also looks like the results might differ based on the data error that you request (see the first post you linked to). I never got a reply, so could you check if this makes a difference?
It seems logical to me, because Poisson errors cannot be computed for non-integer quantities. Therefore, when you scale bins, you need a different error model.

thank you for the fast reply, due to the errors that is the reason why I wanted to avoid scaling the Histogram but I will try what you posted and let you know.
Another solution would be if it’s possible to have the PDF normalized in different regions of the same RooPlot. Like if I have n bins of a given size, have the PDF normalized to those bin widths and then if I have m bins of another size normalize it to those accordingly. I guess the PDF will not look great but at least it will be the correct representation of it. Can this happen ?

You will, however, have to compute the “target normalisation” by hand, i.e. sum over the bins in the histogram, and multiply/divide this value by the bin width.