Show HN: Fursona Editor

But 'just another generator' is a very "Dropbox is just rsync" sort of middlebrow HN dismissal. It takes a lot to make a model and site like TFDNE.

Here's just a very brief summary, skipping over lots of problems along the way. To make it, Arfa had to scrape all of e621. Then he had to hand-annotate several thousand images with the eyes/mouth/face bounding boxes to train a YOLO NN to detect faces, allowing him to crop them out of e621 as a whole. Because he did facial landmarks, he could align the eyes/mouth to the center of the image like FFHQ, throwing out weird faces. This data cleaning gives much higher face quality. Then with the cropped faces, he trained StyleGAN 2 to convergence on a TPU pod acquired through TFRC. (The StyleGAN 2 codebase does not remotely support TPUs, or TPU pods, but Shawn Presser had spent a month adding support and debugging it and setting up infrastructure letting Arfa iterate on S2.) After experimenting with various levels of data cleaning and attempting conditional StyleGANs, he picked a final model, and generated 100k images. These images were then upscaled with waifu2x (itself increasingly a challenge to install since it uses Torch - the original, not PyTorch). Also, about 10k were generated with different psi settings to create 'weird' faces, since we knew from TWDNE how very much people enjoy weird and failed faces. Then a Hetzner dedicated server had to be set up to host it (GCP cloud bandwidth being ruinously more expensive by a factor of ~100). Arfa reused These Waifus Do Not Exist, but worked with the original maintainer to improve scroll control and optimize depiction. That just gets you a static website, of course. But once launched, he had to deal with an onslaught of criticisms and falsified DMCAs or accusations of plagiarism. Fortunately, that's over. But how do you edit? You can't just use a raw StyleGAN model. To enable editing even in a Colab notebook, you have to figure out what combinations of latent vectors do what. That requires setting up Ganspace and exploring hundreds of combinations that Ganspace deduces, in order to find interesting ones and play with them until you figure out what it 'means' and can give it a human-meaningful label. Then all of this has to be cleaned up, hosted, and turned into a usable Colab notebook. But wait, there's more! Apparently Google Colab will just disable notebooks for no apparent reason! So now he has to change the domain to redirect to a new (working?) notebook... And so on.