@alexia i kinda disagree, given the things that AI really accels at are also really horrible (facial recognition for mass surveillance, giving it a giant dataset of an entire social site and asking "where give me everyone who said something like .. " and so on. "automatically point the gun at the person and shoot them + now with image recognition!" and other such 'tasks' which they are currently making ai for the express purpose of "solving" ..

the fact that in general the AI "dream" is to create slave labour where they can't fight back; and all the people going off about "safety" is how they might try not let you do that, while going off about much "benefits to humanity" when, such "benefits" in practice are, badly written stories and art and .. the things it's "good at" are just 'can hurt people easier'

in general, i think these are much bigger issues and in general indicate a issue with the technology as a whole, that is much much larger than just unconsentual training data.

these last three are open source according to the osi definition, meaning you can get access to their training data and train them yourself if you have enough money and technical knowledge.
This is usually a big green flag that the model has not been trained on copyrighted or otherwise unconscenting data, because people would be able to tell much, much more easily, as you can just do a simple search in the publicly available datasets, as well as compare the final performance of a model you trained yourself that there is no unknown extra data involved.

As of yet i do not know any diffusion model trained this way (sadly), so i just have to trust that freepik did not lie about training F-lite solely on data they own