AI Image Extender Expand an image to any shape — nothing gets cropped

These two get confused constantly, and picking the wrong one wastes credits, so it is worth one paragraph. Upscaling adds PIXELS: the same picture, same framing, same content, rendered at a higher resolution so it survives being printed or viewed large. Extending adds PICTURE: the resolution stays in the same league, but there is now more scene than the camera ever recorded. If your image is too small, upscale it — that is a flat one-credit pass on a different page. If your image is the wrong shape, extend it here. Doing the second when you needed the first is the most common mistake in this category.

How it works underneath matters too, because it explains the page’s one real caveat. Your photo is not sent to the model as a photo. The browser builds a canvas at the shape you picked, drops your image into the middle of it, and builds a second black-and-transparent image that tells the model which pixels to keep and which to invent. Those two go up together, and the model paints into the transparent area using everything it can see in the kept part — the light direction, the surfaces, the perspective, the colour of the day.

The caveat is that the model returns a NEW rendering of the whole frame, not your original with a border glued on. In practice the kept area comes back looking like itself, but it is not pixel-for-pixel identical — fine detail can shift very slightly. This is inherent to how the outpaint pipeline works, including in the main Renoise editor, so keep your original file rather than treating the extended version as a replacement. For the same reason, the output resolution follows the model’s own tier rather than your canvas: in our tests a 1536×864 composite came back as 1280×720.

One last practical note, and it is a slightly counter-intuitive one: be generous with the extension. A tiny expansion gives the model a thin strip to reason about and it can produce mushy filler; a large one gives it room to construct something coherent. Our most convincing result was also the most ambitious — a 3:2 street photo pushed all the way to 9:16, where it built an entire upper storey and a cobbled street that were never in the frame.

Most tools on this site give you five or six free runs, and one gives twenty. This one gives two, and rather than bury that we will say why.

Outpainting needs a model that accepts a mask — an instruction saying "keep exactly these pixels, invent everything else". We tested the cheaper options before settling. Nano Banana 2 Lite, which powers several other tools here at four credits a run, rejects a mask outright. Nano Banana 2 accepted the request and returned something worse than a failure: our original image letterboxed onto white, with the furniture quietly restyled — it had ignored the mask and simply run a normal image-to-image pass. A tool that looks like it worked and did not is the one genuinely bad outcome.

So this page runs GPT Image 2, which does honour the mask, and which costs eight credits a run. Twenty free credits divided by eight is two. We would rather give you two runs of something that actually extends your photo than six runs of something that pads it with white and calls it done.

That also shapes how to use the free ones. Pick the shape you actually need rather than experimenting, keep the default instruction unless something specific needs saying, and use a photo with a continuous surface at the edge if you have a choice — those three things get you the most out of two attempts.