☁️BizyAir Load GroundingDino Model

GroundingDINO is the model that turns plain-English object descriptions - "the red mug," "the dog's collar," "the car in the background" - into a bounding box on your image. No clicking, no training, just text in and a box out. This node is step one of the standard pairing: load the detector here, then feed it (alongside a SAM predictor) into BizyAir_GroundingDinoSAMSegment to go straight from a text prompt to a pixel-accurate mask.

How it works

Per the wider ecosystem's standard workflow, GroundingDINO finds the object and SAM turns that box into a precise mask - this node only handles the first half, and it does nothing visible on its own. Right now there's exactly one weight on offer: GroundingDINO_SwinT_OGC (694MB), the original standard checkpoint from IDEA-Research's open-vocabulary detector, so the model_name dropdown isn't really a decision yet, just a formality that leaves room for future additions.

This kind of text-grounded detection is the backbone of a specific class of workflow: anywhere you'd otherwise have to hand-paint a mask around something you can describe in a sentence. Product photography pipelines lean on it to isolate a product without touching its edges by hand; selective-editing workflows use it to target "the necklace" or "the left hand" for inpainting without a mask editor open at all. This loader is just the entry point - the model itself is doing nothing until you pass its output into an actual segmentation node.

The inputs and outputs that matter

• model_name - an enum with one current choice, GroundingDINO_SwinT_OGC (694MB).

Output: GROUNDING_DINO_MODEL - wire it into a downstream detector node (BizyAir_GroundingDinoSAMSegment in this pack).

How to install it

ComfyUI Manager: search BizyAir, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir

then restart. BizyAir wants an account and API key set up on first use per the README.

Common issues & troubleshooting

Loading it repeatedly inside the same workflow. Load it once at the top of the graph and reuse the GROUNDING_DINO_MODEL output across every segmentation call downstream - there's no reason to add a second loader node for a second object query in the same run.

Wondering why there's only one model choice. That's accurate as of this pack's current release, not a display bug - SwinT_OGC is the standard GroundingDINO checkpoint most of the ecosystem ships with, and there's nothing to configure beyond picking it.

Not sure what to plug the output into. This node produces the detector only - it does nothing on its own. Pair it with a SAM predictor loader and BizyAir_GroundingDinoSAMSegment to actually get a mask out.