
Objective: Use Runway to train my own StyleGAN2 model using butterfly images from Flicker.
I first scraped 1919 images from Flicker using the keyword “butterfly”. Then, I uploaded the dataset to Runway. I used StyleGAN2 with the Objects pre-trained model and 3000 steps to train my butterfly model. I applied center crop to all my images to make the dataset more uniform. After 4 hours, my model produced a FID 92.87. I suspect the high value is due to limited number of steps and the complexity of the dataset. Pictures in the dataset varied since butterflies were in various environments such as leaves, flowers, or trees. The butterfly wings were also in different positions from picture to picture.
The final output was not the best. Many of the images reflected the background environment of the butterfly dataset rather than the butterfly themselves. I am curious how the results would have differed if I had used faces or illustrations as the base pre-trained model instead of objects.
Here are my favorite output images where the butterfly wing structure is very apparent.