What the model is actually doing
An outpainting pass starts by placing your image on a larger canvas. The empty area becomes a mask, and the model is asked to fill that mask so it agrees with the pixels it can already see. The original image is not regenerated. Those edge pixels are the only evidence the model has about what belongs outside the frame.
That one fact explains most outpainting behavior. If the edge of your photo holds half a doorway, you get a doorway. If it holds a shoulder, you very often get an entire second person, because a shoulder is the strongest hint available and the model completes the pattern it recognizes. Tools label the feature differently (generative fill, uncrop, AI outpainting) but the underlying mask-and-fill step is the same everywhere.
How far to extend per pass
Extend 25 to 40 percent of the canvas at a time. Three passes at 30 percent beat one pass at 100 percent, and each pass gives you a checkpoint you can reject before compounding an error.
Two other habits matter as much as the step size:
- Keep real overlap. The mask should begin a few dozen pixels inside the existing image so the model blends instead of butt-joining. Most tools expose this as feather or mask blur. Wider feather hides seams; too wide and it repaints detail you wanted to keep.
- Watch the pixel budget. A 1024-wide render spread across a canvas twice as wide leaves your subject at half the detail it had. Extend, then upscale.
The failure modes worth recognizing
- Duplicated subject. A person, tree, or window repeats in the new area. Cause: subject too close to the edge. Fix: smaller expansion, tighter starting crop, negative prompt.
- Horizon drift. New ground and sky do not line up with the old. Cause: a narrow strip carries almost no perspective information. Fix: extend one side at a time and state where the horizon sits.
- Lighting mismatch. The new region is lit from the wrong side. Fix: name the light direction in the prompt, since the model will not infer it reliably from a strip of shadow.
- Texture cloning. Visibly tiled grass, brick, or bokeh. Fix: reroll with a different seed. Rewording rarely helps with this one.
Writing the prompt for an extension
Describe the whole finished picture, not only the part being added. Conditioning is global, so a prompt that mentions the new area alone competes with the pixels the model can already see, and you get a scene that argues with itself.
Then state the frame you want. Outpainting is nearly always a reframing job, so say what belongs in the added strips: empty wall, more sky, a second doorway, nothing. Saying "nothing" out loud is often what stops the model from inventing a crowd.