What the AI does, and what it doesn't
AI fill uses MI-GAN, an inpainting network that takes a 512 × 512 square with a hole and paints what belongs there. To extend a photo, the new canvas is covered in overlapping squares, nearest the photo first; each square sees the photo plus everything already generated, so the far edges are built on top of the near ones.
It's good at continuing scenery, rooms, foliage and soft backgrounds. It is not a text-to-image generator: it can't add an object you describe, and it can't reliably finish a face or a person cut off at the edge.
Honest limits
- The model works on a canvas up to 1024 pixels on its long side, then the result is scaled up. On big prints the AI edges are softer than the photo.
- It works best when you add up to about half the photo's size. Far beyond that, results get vaguer.
- The first use downloads 26.6 MB. After that it's stored in your browser.
- Speed depends on your device: with WebGPU a typical extension takes seconds; on CPU only it can take a minute.
