EVF-SAM

Why we included this project

EVF-SAM is a solid reference implementation for teams that want to segment images by describing the target in natural language instead of clicking or drawing a box. It works by fusing the image and text early in a vision-language encoder and feeding the result directly into SAM's prompt interface, so it skips the usual pipeline where an LLM first proposes candidate boxes. That keeps the model comparatively small and fast; the repository notes a few seconds per image on a T4 GPU. Weights, a Colab notebook, and Hugging Face demos are all released, so you can run your own comparisons on referring-expression benchmarks like RefCOCO without much setup. The zero-shot video segmentation demo built on SAM-2 is also worth checking if your project needs to track objects across frames.