It's useful progress. Decent-fidelity local-scale inference means that you can create a product that generates throwaway images frequently without worrying about cost. Thus far every product I've seen that generates images is metered, which severely limits the value. I don't know if this is actually at the "decent fidelity" point yet.