Seems like every mask is one pass. So even multiple penguins are multiple inferences. But if it continues from SAM 2, the heavy compute is the original image en

Seems like every mask is one pass. So even multiple penguins are multiple inferences. But if it continues from SAM 2, the heavy compute is the original image encoding which is reused and cashed for every inference. No idea what they will do for their API, but from a compute perspective the prompt is free once the image is processed.