Noise figure is a property of receivers, not sources. Signal generators don't have noise figures.

The MAX2870 synthesizer chip used in the project has pretty good phase noise specifications. I don't see any broadband noise above the background in the SDR trace shown in the article. The second harmonic is more than 30 dB down which is not bad at all. The only thing not looking that great is the spur close to 400 MHz.

The noise floor of the measurement receiver is terrible though so you really don’t know what the true performance is here.

I don't see a resolution bandwidth in the plot. How do you judge the noise floor without that?

RBW is fixed at 1.75MHz on the HackRF. But it's the measurement setup as a whole (that is the measurement receiver). The noise floor of this measurement is represented by the visible noise floor at -70dBm. That's a combination of a number of factors. Plus the HackRF has poor dynamic range and only 8 bits sampling which is quite terrible.

Make a better one and submit your link here

You don’t build RF signal generators. You buy them. That’s pretty much it. You don’t have the money or engineering ability to build one at home with any reasonable performance characteristics.

A 30 year old Marconi or HP crate will set you back $500-1000 still today because of the performance requirements of what constitutes an RF signal generator. A random PCB doesn’t have a well shielded step attenuator, known phase noise performance, known frequency stability, known harmonic performance and predictable levelled and attenuated output or reference modulation capability.

Go look at a tear down for a Marconi 2019A, my favourite generator and see what is really involved.

An RF signal generator is for making measurements with. This one isn’t. It’s the equivalent of a vuvuzella at a football match as far as signal sources go.

Agreed. A 45-year-old HP 4650B with its cavity oscillator has fancier physics going on than you're likely to build at home unless you make building signal generators your hobby for the next decade or two.

Just looking into what would be involved in building a stepped attenuator comparable in insertion loss, accuracy, and isolation to the HP-335 series from 1958 should give one pause.

The Marconi looks nice.

In my experience, modern integrated synthesizer or DDS chips perform not that much worse than old benchtop signal generators. You don't get creature comforts, such as calibrated electronic step attenuators, and may have to calculate the frequency divider settings by hand. For many hobby or research projects they are certainly good enough.