Outpainting: Extending an Image Without Breaking It

Outpainting asks the model to invent what was outside the frame. It is a different job from generating a picture, because most of the answer already exists — the new canvas has to agree with the light, the perspective and the material of what is already there. When outpainting fails, it almost always fails on one of those three, and the prompt is usually the reason.

Describe the New Area, Not the Old One

The most common mistake is re-pasting the original prompt into the extension. If the source image was “a woman in a red coat on a station platform” and you feed that back while extending to the left, you will very often get a second woman in a red coat. The model reads the prompt as a description of what should be in the region it is painting, and the region it is painting is empty.

Write instead what the new strip should contain: more platform, an empty bench, a departure board in the distance, sky above the roofline. Keep the same medium and lighting words as the original — those apply to the whole picture — and drop every subject word that names something already in the frame. Duplicated subjects disappear the moment you stop asking for them.

Extend in Steps, and Keep the Overlap Wide

Trying to add half a canvas in one pass gives the model very little to condition on at the far edge, and that is where drift shows up: the colour temperature slides, the horizon tilts, the wall changes brick. Extending in narrower increments, each one overlapping a generous band of finished image, keeps every new pixel close to something real.

The overlap band matters more than the increment size. A thin sliver of context produces a visible seam, a hard line where the tone steps. A wide band lets the model resolve the gradient across it. If a seam appears anyway, do not try to prompt it away — run another pass centred on the seam itself so the model repaints across the line rather than up to it.

What to Do About Perspective

Perspective is the part that no amount of prompt wording reliably fixes, because the model is extrapolating vanishing lines it never explicitly computed. Straight architectural elements are the giveaway: a ceiling that bends, a floor tile grid that changes direction halfway out.

Two things help. Naming the camera position in the prompt — eye level, looking slightly down, wide shot — gives the extension something consistent to aim at. And choosing where to extend matters: extending into sky, water, foliage, fog or shallow depth of field is nearly always clean, while extending along a strong receding line is the hardest case. If you know in advance that a picture will need a wider crop, it is far cheaper to generate it wide from the start than to rebuild the geometry afterwards.

When an extension is close but a small region is wrong, treat it as an inpainting problem rather than redoing the whole strip. Outpainting sets up the new space; local repair finishes it.