Grounding Model (down)Loader

This node is the pack's whole pitch in one dropdown: 19 bounding-box detection models - GroundingDINO, MM-GroundingDINO, OWLv2, Florence-2, and YOLO-World - with zero install per model and a cache that makes switching between them nearly free. It loads whatever you pick, auto-downloads the weights on first use, and hands a ready-to-go GROUNDING_MODEL to the Grounding Detector node.

Why would one person need 19 detectors? Mostly to stop reinstalling node packs every time a better model drops. The author's stated reason for building the pack was chasing down failure rates in background-removal pipelines - trying every grounding/SAM2 combination until one nails the object. With a single loader, you can A/B five models in one session without rewiring anything.

Which one to pick

• Florence-2: Base is the default, and it's a sensible default: small (0.23B params), fast, and competent at descriptive prompts. If you just want boxes, start here.
• GroundingDINO SwinT is the classic lightweight choice; SwinB is heavier but more accurate. Both are ~700–940MB.
• MM-GroundingDINO models print their mAP right in the dropdown - the "Large All Datasets (60.3 mAP)" is the accuracy king here if you can afford the VRAM and the download.
• YOLO-World v8s/m/x are the speed picks, trading some semantic understanding for near-real-time detection.

The honest framing: bigger name almost always means bigger VRAM footprint and a slower first download. There's no reason to start with OWLv2 Large when Florence-2 Base does the job.

The inputs

Just two. model (the dropdown, with keep_model_loaded as an optional toggle that defaults to true). When it's on, the model parks in VRAM so every subsequent detection is instant - the pack's smart cache keys each model in memory, so even switching models mid-workflow is cheap. Turn it off if you're sharing the GPU with something heavy like Flux; the README's advice is blunt: it offloads to CPU after each detection, slower but far more memory-friendly.

Output is the single model wire, which feeds GroundingDetector. That's it - this node doesn't detect anything itself.

Installing it

Same pack install as everything else in ComfyUI-Grounding:

cd ComfyUI/custom_nodes/
git clone https://github.com/PozzettiAndrea/ComfyUI-Grounding
cd ComfyUI-Grounding
pip install -r requirements.txt

Then restart ComfyUI (or install via Manager by searching "Grounding"). The first time you select a model it downloads from Hugging Face - expect a wait, and expect the bigger MM-GroundingDINO / OWLv2 large variants to be the slow ones. YOLO-World weights come from Ultralytics' own release assets instead.

Where people get burned

• Dependency weight - the pack pulls in transformers, timm, einops, and ultralytics. Most are already in a modern ComfyUI, but if you see import errors in the console on startup, it's usually a broken version of one of these.
• The Ultralytics history - YOLO-World's dependency had a well-known December 2024 supply-chain compromise that reached ComfyUI users through Impact Pack. Buy your ultralytics from PyPI, keep it pinned, and don't chase random prebuilt wheels.
• First-run confusion - the dropdown shows all 19 models, but nothing downloads until you actually select and run one. A quick "why is it hanging?" during first run is just the model downloading; it's cached after that.