Prompt detection

When inpainting, it detects objects that were in the prompt, but sometimes it detects objects that weren’t. How do I limit the area of detection in an image?

The issue does not name a source file, test, or entry point. Start by reproducing the inpainting prompt-detection behavior in the repository’s example inference notebooks, then determine how the detection area is currently selected and define what limiting it should mean. Done means detections are restricted to the intended image area without including unrelated prompted objects.

Written by the indexing model from the issue text.