Yes, the problem with telling the model to correct something or avoid making a certain mistake often isn’t about following instructions. If the errors are caused by limitations of the model itself, they can’t simply be fixed with a prompt. It’s like telling a person, “Be better at mathematics.” If they don’t know mathematics, the instruction alone won’t help.
What you’re actually doing when you ask the model to improve something is directing more of its attention to that particular aspect. Ironically, this can sometimes lead to overcompensation, causing new problems to appear elsewhere in the image.
One technique that can help is repeating the most important parts of a prompt, either verbatim or with only slight variations. Since the order of the prompt also influences the result, you can even repeat those key instructions near both the beginning and the end of the prompt.
These systems have become remarkably efficient in a very short time, but they are still not perfect. It’s similar to many engineering problems, the last 10% of quality often requires as much time and effort as the first 90% that builds the basic foundation. This is especially true for photorealistic images, where people have almost no tolerance for errors. Achieving those final 10% of realism requires a great deal of additional development.
From what has been officially said about version 2.0, one of the main goals was to make the model faster and more efficient, allowing it to generate images with less computational power. In some situations, that optimization may come at the cost of image quality-particularly in the cases where those final 10% make all the difference.
Topology is especially challenging because it requires an understanding of the image as a whole rather than just its individual parts. That kind of global reasoning is exactly where neural networks become computationally expensive and resource-intensive.
