Comment two stage design (Score 3, Insightful) 34
My understanding is that these are all done in multiple stages.
The big network doesnt have artificial bias or artificial limitations, but thats the second stage.
But thats the second stage. The first stage, the small network, or maybe just an expert system, is transforming your prompt into a network-ready vector, and this is where they impose limitations and biases.
You ask for "images of a family" and the first stage transforms your query into "[preamble of rules.] Generate images of a diverse family"
Its going to stay this way because the time and dollar cost of biasing and limiting the big network is prohibitive. Aint nobody got time to label the training set. Also, it allows offering differing levels of modification based on the customer. The general public version wont even be the most limited of them, and very special customers dont have limitations or biases injected.
Now, in the case of an expert system for stage 1, thats going to be easily tricked unless its very heavy handed on the limitations.
The big network doesnt have artificial bias or artificial limitations, but thats the second stage.
But thats the second stage. The first stage, the small network, or maybe just an expert system, is transforming your prompt into a network-ready vector, and this is where they impose limitations and biases.
You ask for "images of a family" and the first stage transforms your query into "[preamble of rules.] Generate images of a diverse family"
Its going to stay this way because the time and dollar cost of biasing and limiting the big network is prohibitive. Aint nobody got time to label the training set. Also, it allows offering differing levels of modification based on the customer. The general public version wont even be the most limited of them, and very special customers dont have limitations or biases injected.
Now, in the case of an expert system for stage 1, thats going to be easily tricked unless its very heavy handed on the limitations.