You upload up to ten photos and write what you want. "Put a red bandana around the dog's neck." "Print the dog from the second photo onto the mug in the first." It returns one new image built from your references. With no photos, it works as a regular image generator, and it can letter short text like a poster title into the picture.
The difference is placement. You draw a rectangle on the canvas for each element: a sticker in the top right corner, a window on the left wall, a figure moved further up a hill. Then you move, recolor, replace or remove an element by pointing at its box. Boxes guide position and size but are not a hard cutout, so an object can spill slightly past its edges.
It comes from Black Forest Labs, the team behind the FLUX image models. A separate helper model first expands your short request into the detailed scene description the image model needs, and the demo runs that step for you.