Many image generators produce images in a fixed order, left to right and top to bottom, like reading a page. That works for creating a whole image, but it cannot handle tasks like filling in a missing part, extending the edges, or editing a region, because those require generating from any starting point, not just the top-left corner. Retraining such a model from scratch to work in any order costs roughly three times the compute, and it loses the ability to use fast, existing inference tools.
Researchers found that an already trained model already knows what images look like; it just never learned to express that knowledge in any order. They added two small changes: positional information that tells the model which location it is predicting next, and a training schedule that gradually shifts from the fixed order to random orders. This gently repurposes the existing model instead of starting over.
On ImageNet, the adapted model generates better inpainting results than a common baseline (FID 6.21 versus 10.44, where lower is better), and it cost only a fraction of the from-scratch training budget. The same weights now support inpainting, outpainting, and editing without extra machinery, while still working with the standard autoregressive serving stack.