This project is an exploration into playing with watercolour imagery using Machine Learning. I have no idea where this will end up, but here goes....
I started off by getting almost 2000 watercolour paintings. I fed them into StyleGAN2 using RunwayML and generated these beautiful abstract watercolour artworks.
I asked the AI what does it see in the abstract watercolour paintings it created (using annotation algos) . Interestingly enough it said that it sees some forms in the abstracts, which I previously never saw. But when read them, I did see some of them too. I think I might have something going on in this dance with the AI.
What I did next was to feed the text back to the machine using AttnGAN to find out "is that what you are really seeing machine? Or do you have a preconceived notion in mind". The machine did have a similar yet different ideas about them. I love this dance!
My friend and Georgia Tech Professor Devi Parikh used some of the abstracts to have a conversation with a bot to find out what exactly it sees. Check these out:
I tried feeding one of the watercolour abstracts as a seed image and then running a loop image-text-image-text. Some of them generated interesting results after sometime:
While others went in an infinite loop
I started generating some interesting artworks out of the system, by cherry picking what looked best
Making some beautiful rocks in Watercolour
Up next I used one of the watercolour abstracts as a seed to redraw it with my own human imagination. No drastic changes though.
Then, I tried using the watercolour abstracts in a 3D inpainting algorithm. Here are few results:
One of the reasons I cherry pick images is because this was trained on the pre-trained faces dataset. So there is an amount of noise in some images. I hope to get more GPU capacity to get better results. But as of now, even these noisy images are interstinig.
I reverted to a model in the earlier stages of the transfer learning cycle and got these interesting Faces + Watercolour mashups.
Midnight Falls - yet another series of abstracts. (Or are they new fashion jeans?)
And finally a deep swim in the latent waters of colors: