visinf/INSID3 — reverse-engineered prompt

Build me a Python project that recreates INSID3, a training free in context image segmentation tool built on a frozen DINOv3 backbone. I want to be able to give it a reference image plus a mask, then a new target image, and have it predict the matching object or part in the target and save a nice visualization.

Please make it usable both as a simple script and as a small programmatic API, with a minimal example that works on sample images. It should support the different DINOv3 model sizes, let me trade speed for quality with image size, and optionally refine masks with CRF if that dependency is available. Also include the evaluation scripts for the segmentation datasets mentioned in the paper, and the semantic correspondence comparison between original and debiased features.

Make setup straightforward, document where the pretrained DINOv3 weights go, and include clear instructions for running everything locally or in a notebook. If anything is unclear, look up the current DINOv3 docs online and wire it up in the simplest working way.