VercelAiSdkVisionCoordinateSpace

Coordinate space the model is asked to report regions in.

- "pixels" (default): absolute pixel coordinates of the transmitted
image. The image dimensions are stated in the prompt. Matches models
trained on absolute coordinates (e.g. the Qwen2.5-VL family).
- "normalized": coordinates relative to the image dimensions, in the
range [0, coordinateScale] (default scale: 1000; left/width
relative to the image width, top/height to the image height). The
finder converts them back to pixels deterministically. Matches the
grounding convention of e.g. Gemma and Qwen3-VL, and is invariant under
uniform image rescaling by the serving runtime.

Model families differ in their trained convention, so benchmark both
settings for a given model (e.g. with @nut-tree/vision-evals).