Author here. The benchmarking tool I used for measuring response size was vegeta, which ignores HTTP headers in its measurements. I believe the difference in size is indeed in the QR code images themselves.