How Does It Work?

The training dataset consisted of ~60k SFW images from e621.net and FurAffinity, cropped and aligned to faces using a custom YOLOv3 network.

As the images are generated by an AI, they are non-copyrightable. I claim no legal ownership or rights to any of the images generated by this AI. Please use them responsibly.

• The network was trained for 33 days on a TPUv3-32 pod.
• The final images were scaled and denoised using Waifu2x.
• TPUs provided by Google's TensorFlow Research Cloud (TFRC) program.