Grounding DINO

Open-set detection went mainstream in 2023 with Grounding DINO from IDEA Research: hand it an image plus a short phrase and it draws boxes around whatever the words describe, with no class-specific training. Paired with Segment Anything as Grounded SAM, it made text-prompted masks easy and became a common auto-labeling engine for building training sets. The official repository has been quiet since August 2024, when its final commit linked to Grounded SAM 2; its only tagged release dates from April 2023, the training code was never published, and the custom CUDA extension still breaks many installs. The model itself holds up: the tiny and base checkpoints remain on Hugging Face under Apache-2.0, and the Hugging Face Transformers port runs them without the fragile compile step. Newer IDEA models in the series, starting with Grounding DINO 1.5, are offered only through a hosted API. For new projects, use the Transformers port, Grounded SAM 2 when you need tracking across video frames, or Meta's SAM 3 when you want concept prompts that return masks directly, after checking its custom SAM License.