LLMDet

LLMDet (CVPR 2025) is a Grounding-DINO-style detector trained with captions generated by a large language model. The default Transformers build is small and quick; the original MMDetection checkpoints are available as extra tags.

Object DetectioniSEE LabApache-2.0

Recommended4/10

VRAM1.4–5 GBweights, plus ~0.5 GB reserve

Tools1across 1 annotation type

Runs onmacOSCPU2 of 4 runtime profiles

Install

boat model install llmdet

Without a tag BOAT installs llmdet:transformers-large. Tags install side by side, so you can keep several.Download: 1.4 GB of weights.

Tools

What LLMDet adds to the annotation editor. Each example loops what you do and what comes back.

Text to Boxes

Text promptBounding boxes

Object Detection

Type a class name such as “car” and press Enter. BOAT previews every matching bounding box before you apply them.

Pick candidates

Tag	Weights	Platforms
llmdet:transformers-largedefault	1.4 GB · 6 files	macOSCPU
llmdet:mmdet-tiny	4.5 GB · 15 files	macOSCPU
llmdet:mmdet-tiny-p5	4.5 GB · 15 files	macOSCPU
llmdet:mmdet-base	4.7 GB · 15 files	macOSCPU
llmdet:mmdet-large	5 GB · 15 files	macOSCPU

Platforms

BOAT picks the runtime profile that matches your machine. Only that profile's code is downloaded.

macOS · Metal

Apple silicon (MPS)

Python 3.11 · 3.2 MB

Linux · CUDA 12.8

NVIDIA, incl. Blackwell

Not supported

Linux · CUDA 12.6

NVIDIA GPUs

Not supported

Linux · CPU

x86_64 CPU

Python 3.11 · 3.2 MB

License

BOAT's adapter code is Apache-2.0. The model's own code and weights keep their upstream license:Apache-2.0.

Bundled license files: BOAT-APACHE-2.0.txt, NOTICE.md, UPSTREAM-LICENSE.txt

Package details

Package: llmdet
Version: 0.0.1
Type: Common model
BOAT Core: ≥ 0.1.0 · < 1.0.0
Idle unload: 300s (change with boat model setup)
Inference timeout: 600s
Manifest SHA-256: b174538d9c50336136ac7104d13b00991f79c63012a3996a190be40a5bedc124
Manifest: manifests/llmdet/0.0.1-b174538d9c503361.json

Upstream notice

BOAT adapter code is Apache-2.0. Upstream model code and weights keep their own licenses.

Model: llmdet
Pinned weight provenance is recorded per variant in ../weights/<tag>/weights.json.
The original MMDetection/LLaVA runtime is vendored from
https://github.com/iSEE-Laboratory/LLMDet.git at commit
53366243fba758ad3d7a042c0c1457dee91a02b2. Each platform copy records this in
models/<platform>/vendor/llmdet_upstream/UPSTREAM_REVISION. BOAT adds local
weight-path resolution, portable CPU/MPS operators, and standard result conversion.
Runtime dependencies are installed from pinned distributions with their own license files.

Upstream license: https://raw.githubusercontent.com/iSEE-Laboratory/LLMDet/main/LICENSE
Model card: https://huggingface.co/iSEE-Laboratory/llmdet_large