lightning-sam

Why we included this project

Stock Segment-Anything handles some images poorly, and the author's example is text on documents. This library adapts SAM to your own data by fine-tuning it on any COCO-format dataset, using bounding boxes as prompts, and it runs on Lightning Fabric. The README is honest that this is an experiment, a proof of concept to see whether box-prompt fine-tuning improves mask quality. It supports only box prompts and includes training and validation loops, so treat it as a reference implementation rather than a production pipeline. The reported IoU gains on TrashCan and WGISD give a concrete sense of what fine-tuning can buy you, which helps you judge whether the effort is worth it for your own domain.