-Using her name to describe/advertise the fine-tuned model.
-Using her illustrations to fine-tune the model.
-Using a larger body of potentially unlicensed images to train the base model.
For the first, if we had decided that the other steps were fair use or whatever, would it be better or worse if the fine-tuned model had been made available with no mention of the identity of the author of the training images? I'm not sure.
For the second, there is surely a limit after which this sort of thing becomes unambiguously unacceptable. Suppose you fine-tune so aggressively on a small dataset that eventually the model simply reproduces exactly the training images. Now you're obviously violating copyright. But where exactly is line before that? If I have a base model that was trained on fully licensed images, and I make one single gradient descent step using a copyrighted image, making imperceptible changes to the model's output, surely the resulting images not suddenly in violation. It seems to me that the standard should be that if a human were to draw the output by hand after looking at the training images, would we consider it a violation of copyright? As a thought experiment, imagine someone who lacks the ability to draw but can instead hand-write the weights of a neural network to produce the desired output - it shouldn't matter which process they use.
For the third, what if I spent a long time prompt engineering on a model trained entirely on properly licensed data and was able to generate a prompt format that produced the outputs we see from this fine-tuned model? In other words, for any generative model, there is a space of reachable outputs, and it's not so clear that these images did not already lie in that space before fine-tuning.