From what I can tell, it doesn't look like the recent GPT-4o image generation includes the research of the NeurIPS paper you cited. If it did, we wouldn't see a line-by-line generation of the image, which we do currently in GPT-4o, but rather a decoding similar to progressive JPEG.

Going off my bad memory, but I think I remember a comment saying the line-by-line generation was just a visual effect.