by Mary Kate Skitka
Cantaloupe Dadaism attempts to create fashion from non-traditional sources of inspiration. Most of the images used to generate the looks were from inanimate objects but also include animal and people source images. More broadly the project begins to explore how AI can contribute to new fashion trends and stretch our imagination of how we may express ourselves.
Going into this idea I had originally hoped to generate more realistic looking clothing articles, especially with more texture than what I came out with, however I was still pleased with the output and actually enjoyed its abstract nature as it allowed me to interpret the clothing which added an additional creative element. Getting the model to output images and styles that were not primarily one color or that looked human was difficult. To address this I mainly used a trial and error method. I probably ran through 50 images and I ended up using 20 in the final look book.
On exploring additional models and post editing:
In order to make the images more realistic I did play around with a couple of solutions. I tried feeding the output of the fluid warping GANs into various image generators and colorers but nothing achieved what I was looking for (See examples of this below). Eventually I did take video of various stances and positions and then re-fed the first output to the fluid warping GANs a second time. For certain cases this helped create clearer poses and added clothing features such as legs, skin pigments or shirt outlines. Additionally I did some post-process cropping in photoshop on some of the images but I did not change the integrity of the photo manually.
Original output image from CD GANS
Post processing using BigBiGAN model in Runway, probably the closest model I found to create further realism.
Original output image from CD GANS
Post processing using style2paint model in Runway
Post processing using SPADE-COCO model in Runway
I wouldn't do anything differently I really like how everything turned out however I think there is more that could be done to the images. I would love to find either a model to add texture or manually add more texture. I also think that adding faces could be interesting. Also my teacher gave me the idea that creating these garments physically could be very cool.
I would say my main take aways from this assignment were growing a familiarity with Runway and specifically getting to know the Cataloupe Dadaism model and it's idiosyncrasies. For example I noticed that the model tended to use the colors from the center of the reference image to overlay onto the human mannequin. I also got more comfortable thinking about combining models and how to do that in Runway.
