Promptable Image Segmentation

Promptable image segmentation studies how a model can turn a user's or another model's cue—such as a click, box, scribble, text description, or example mask—into a precise region. Much work builds on Segment Anything models, adapting them with lightweight adapters or fine-tuning, while other methods generate better prompts automatically or use retrieval and visual examples to guide frozen models. These approaches support interactive correction and open-vocabulary segmentation across settings such as medical imaging, remote sensing, microscopy, and video. A central challenge is that results can be sensitive to prompt choice and domain shift, so evaluations increasingly test realistic human prompts, robustness, uncertainty, and the trade-off between accuracy, interaction effort, and speed.