Generate two images, get a vision model to judge, then RL reward the better one. And yes, OpenAI employees on twitter bragged about the model's SVG capabilities. So there are clearly people working there who care about it.