# Queue several runs and pick the best
Generate several variations at once instead of re-running a single generation until it works.
Image-model output varies from run to run, even with a fixed layout and fixed references. A single re-roll loop costs time per attempt and gives no basis for comparison. A parallel batch shows the spread of results side by side, so the best candidate is chosen from a set. In the demo it is the author's default for the final scene, where details such as the added polar bear or the touching bottles do not always land on the first run (see Multi-mannequin composition).