![]() If doing training, I'd highly reccomend using a GPU. ![]() The dataset is available on Google Drive link. ![]() ![]() Credits to who built the map generator I used to construct the dataset. I used the Selenium IDE to build the dataset of fantasy maps. Results coming soon! Currently training a 128x128 version of the net, and things are looking good. Conversely, we can compute discriminator losses from disc_image(x), disc_image(r), disc_image(r') and disc_latent(z), disc_image(z'), disc_image(z''). We can compute reconstruction losses between recon_loss(x, r') and generator GAN losses from disc_image(r), disc_image(r'), disc_latent(z'), and disc_latent(z''). In this case x and r are images, z are latents. For more details, definitely check out that repo! I'm working on having this GAN though scalable to larger sizes (originally 96x96 - want to get up to 256x256.)īelow is a rough architecture flow diagram of what's happening during training. Implementation details: A simpler version (i.e., removal of the conditional image generation functionality) of the AE-GAN I implemented in my Pokemon Sprite Generator project. Using Generative Adversarial Networks (GANs) to produce awesome looking fantasy maps.
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |