ImageText Remover: box out the words, rebuild the picture

Field notes

The workflow is deliberately box-based: users can paste a screenshot, nudge a region by one pixel or ten with Shift-arrow, review region thumbnails, and exclude individual boxes before running. A fast diffusion fill requires no model download, while the structure-aware option fetches a model after showing its byte count. Batch mode applies one box set across same-sized frames, and exports retain the incoming format as the first option.