LoRa trained on various screenshots of flash games. For some reason draws eyes wide, although there weren't many images in dataset like that. With character also learned style, if you want to reduce it's influence — use flat color and photo background in negative.
