But it is still preferable to train on synthetic images than on the original works, don't you agree?

When the artist refuses to allow their images be used for training AI models, or it is impossible to get permission for other reasons such as not knowing the correct contact information, if the AI uses variations it won't learn to imitate the originals closely. Variations should be OK because they have no copyright, as the courts decided. Seems like a better compromise than either indiscriminate training or making AI impossible to train.