Zoom to Inpaint lets you inpaint in higher fidelity. With normal inpainting, the whole image is regenerated but only the masked area actually changes — so if your mask is small, only a little of the model’s attention goes to it.
With Zoom to Inpaint enabled, the start image is “zoomed in” so only a small area around the mask is regenerated, at the full resolution of the generation. That means much more of the model’s attention is focused on the area you’re changing.
Because only the area around the mask is generated, your prompt should describe only the contents of the masked area — not the rest of the image.
With Zoom to Inpaint you can also set Mask Padding — the number of pixels added around the mask to give the model context. Higher padding means slightly less zoom (and slightly lower fidelity); lower padding means more zoom but less context, which can make the inpainted area look different from the rest of the image.