Generative fill covers two related operations that solve real, mundane problems.
Inpainting: Filling a Selected Region
You select an area inside the image and the model synthesizes new content for it, matching the surroundings. Uses:
- Removing objects — a bin, a photobomber, a stray cable, a logo you don't have rights to.
- Repairing damage — tears, scratches, missing corners in scanned photos.
- Replacing content — changing a product colour, swapping a background element.
Outpainting: Extending Beyond the Frame
The model generates plausible content *outside* the original borders. This is the one that solves a genuinely common problem: changing aspect ratio without cropping.
A 3:2 photo needs to become a 16:9 banner. Traditionally you crop, losing the top and bottom. With outpainting you extend the sides instead, keeping the entire original composition and letting the model invent more sky, more wall, more floor. For backgrounds this works remarkably well.
Other uses: adding headroom above a subject that was framed too tightly, extending a background so text has somewhere to sit, and recovering a composition where you stood too close.
Where the Seams Show
Both techniques fail in predictable ways:
- Structured content. Extending brickwork, tiling, window grids or text produces misalignments — the model doesn't understand that lines must continue exactly.
- Perspective. Extending a floor or road often bends the perspective, because the model matches texture rather than geometry.
- Faces and bodies. Adding a limb, or extending a person beyond the frame, frequently produces anatomical mistakes. Extend backgrounds, not people.
- Repetition. Large extensions may duplicate a distinctive element — the same tree twice, a repeated cloud shape.
- Colour and noise mismatch. Generated regions are often cleaner and slightly different in tone than the original; on a grainy or heavily compressed photo the patch is visible precisely because it's *too* clean.
Practical Technique
- Extend in small steps rather than doubling the canvas at once. Several modest extensions produce better continuity.
- Give the model context — include a generous margin of the original when selecting, so it has material to match.
- Prompt when the tool allows it. Describing what should be in the extended area ("continuation of a plain grey wall") beats leaving it to guess.
- Match noise afterwards. Add a small amount of matching grain to the generated region so it blends with the original.
- Inspect at 100% along the boundary, which is where errors concentrate.
- Generate several variations and pick, rather than accepting the first result.
The Disclosure Question
Removing a bin from a holiday snap is retouching. Extending a background so a banner fits is production work. But the same tools can:
- Remove a person from a documentary photograph.
- Add content to a news image.
- Extend a product photo to imply something that isn't included.
The line is consequence: if a viewer would make a decision based on what you added or removed, the edit needs disclosure — or shouldn't be made. Journalism, evidence, insurance, real-estate listings and product photography all sit on the restricted side, and some now require provenance metadata to record generative edits.