Given that you have three replicates for each condition you might want to consider differential binding analysis, e.g. diffReps. It will use the replicates to estimate variance and return windows of the genome that have significantly more binding in one sample compared to the other. The input is the mapped reads, and does not require called peaks, which gets round calling peaks based on arbitrary thresholds.

Those are actually not very different, but both are a little low. There can be a number of reasons why they like that. You can start by looking at the total number of reads, alignment rate, and duplication rate. That may explain why one is better than the other. You can also look at some of the actual peaks in a genome browser and see how the different replicates compare.