Draw a rough layout in the tool's canvas or upload a sketch, add a short line of context naming the shapes, and the AI turns it into a finished thumbnail that follows your composition. The sketch does not need to be good; a stick figure is enough, because the tool reads layout, not artistry.

Gen-2 (20 credits) is best for face consistency and photoreal entertainment thumbnails; Gen-2.5 (16 credits) is faster, cheaper and better for graphics and charts. There is no universal winner, so run the same prompt through both models once and keep whichever fits your channel.