Summary: Removing and adding keywords in image prompting changes the generated images in unpredictable ways. Current prompt adherence is too weak to reliably steer the images through keyword editing. Several months back, I undertook a fascinating challenge proposed by Umesh. The task was to
Fascinating study. Thank you. I tried a simpler experiment some time ago. One thing I tried was using Japanese sumi as a reference, Sesshu in particular. I found the results for western art far better than those for non western traditions. I plan to do this for North West Coast art at some point. Of course that will raise a lot of appropriation and copyright questions.
Definitely the amount of training data has an impact. There are hundreds of thousands of photos of Greek-Roman busts on the Internet to learn from, though very few of them are specifically in blue marble as specified in the prompt. On the other hand, there are plenty pf photos of blue marble in general, and blending the two is one thing that current-generation AI does fairly well. You are right that other types of art with fewer examples to learn from will currently fare poorly, though it does seem there is some progress in small-model AI.
Fascinating study. Thank you. I tried a simpler experiment some time ago. One thing I tried was using Japanese sumi as a reference, Sesshu in particular. I found the results for western art far better than those for non western traditions. I plan to do this for North West Coast art at some point. Of course that will raise a lot of appropriation and copyright questions.
Definitely the amount of training data has an impact. There are hundreds of thousands of photos of Greek-Roman busts on the Internet to learn from, though very few of them are specifically in blue marble as specified in the prompt. On the other hand, there are plenty pf photos of blue marble in general, and blending the two is one thing that current-generation AI does fairly well. You are right that other types of art with fewer examples to learn from will currently fare poorly, though it does seem there is some progress in small-model AI.