Skip to content

Limits

Fulla relies on AI models that are good, but not perfect. Here is what to expect, and what to do about it.

The models return at most about 4K pixels on the long side, 2K for OpenAI and Qwen models. A larger selection gets a result enlarged to fit, which looks softer. When that happens, the card says so: “Large area: enlarged 2.1x, may look soft. Try a higher resolution or a smaller selection.”

To keep it sharp, fill a large area in several parts, as when extending a photo band by band. See Resolution.

Models don’t stop exactly at the edge of the area they were given. Fulla clips their result to your selection, with a soft edge, so nothing outside it changes. But an object the model added can be cut where the selection ends: select generously and leave room for shadows.

Now and then a model redraws the whole picture it was sent, not only the area to fill. The result then looks slightly shifted or different along the edge. Fulla places what comes back exactly where it belongs, but it can’t tell yet that the model did this. Generate again, or keep another variation.

Cancel stops Fulla from using the result, and Fulla never sends a new request after it. But a request the AI service has already received may still be billed.

Colours match your document only when Photoshop keeps embedded profiles: see Colours. A result that an AI service sends back as WebP gets no colour profile yet; PNG and JPEG results do. It hasn’t happened with the current models so far.

Fulla sends the whole photo along as a reference, so the model can judge scale and light. Qwen models have no room for it and only see the area to fill with its margin. The number of reference images also depends on the model.

Every image is made by an online AI service: Fulla needs an internet connection and an account with credit. It runs in Photoshop 2026 (27.11 or later) and the Photoshop beta, not in older versions. The Windows installer is for Windows only; on macOS, install the .ccx with the Creative Cloud app.