![]() Armed with an extremely large dataset, I subsequently began working through particularly promising members of the, emphasizing SOTA & open implementations.Among others, / & (failed to get running), & StyleGAN. Thinking perhaps the problem was too-small datasets & I needed to train on all the faces, I began creating the Danbooru2017 version of. Despite many runs on my laptop & a borrowed desktop, DCGAN never got remotely near to the level of the CelebA face samples, typically topping out at reddish blobs before diverging or outright crashing. (I did a lot of from because she has a color-centric design which made it easy to tell if a GAN run was making any progress: blonde-red hair, blue eyes, and red hair ornaments.)It did not work. So when GANs hit, and could do somewhat passable around 2015, along with my, I began experimenting with of, restricting myself to faces of single anime characters where I could easily scrape up 5–10k faces. Generative neural networks, such as GANs, have to generate decent-quality anime faces, despite their great success with photographic imagery such as real human faces.
0 Comments
Leave a Reply. |