SAM3 Text Segmentation

This is the node most people actually came to SAM3 for. Displayed as "SAM3 Text Segmentation," it takes an image plus a text prompt like dog or red car and hands you back a mask of every matching thing in the frame. No clicking, no drawing boxes, no separate GroundingDINO node feeding it - SAM3 does the open-vocabulary detection and the masking in one shot. Type the noun, get the mask.

That's the leap over the old workflow. The classic setup was two models bolted together: GroundingDINO to turn text into bounding boxes, then SAM to turn boxes into precise masks. SAM3 folds both jobs into one model, which is why "text segmentation" is now a single node instead of a chain. Where you'd use it: pull the subject out for inpainting, isolate a product for compositing, build an automated batch pipeline that finds and masks the same object across a hundred images. It's overkill if you just want foreground-vs-background - grab BiRefNet for that - but it's exactly right when you need a specific named thing.

How it works

Under the hood it's grounded segmentation: the text prompt conditions the model to localize matching regions, and SAM3 produces a pixel mask for each. Because it's open-vocabulary, you're not limited to a fixed label set - describe the object in plain words. The confidence_threshold decides how sure the model has to be before it keeps a detection, so it's your main dial for "found nothing" versus "masked half the image."

The inputs and outputs that matter

The three you'll actually touch:

• sam3_model - the SAM3_MODEL from LoadSAM3Model. Required.
• image - the picture to segment.
• text_prompt - what to find, e.g. person, left hand, yellow flower. Keep it a concrete noun phrase.
• confidence_threshold (default 0.2) - the keep/drop bar, 0–1. Low catches more but risks junk; raise it when you're getting false positives.

Two optionals worth knowing: max_detections (default -1 = unlimited) caps how many instances come back, handy when you want just the biggest few; and offload_model (default off) pushes the model off the GPU after the run to reclaim VRAM. There are also optional positive_boxes / negative_boxes inputs (SAM3_BOXES_PROMPT) if you want to constrain the search region with a box on top of the text.

Outputs: masks (MASK) - the thing you feed into inpainting or compositing; visualization (IMAGE) - a preview with the masks drawn on, wire it to a Preview Image to sanity-check; and boxes / scores (both STRING, JSON) - the detected boxes and their confidence numbers for debugging or downstream logic.

How to install it

It comes with the pack. ComfyUI Manager: search SAM3, install the highest version (nightly if the release misbehaves). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-SAM3.git
cd ComfyUI-SAM3
pip install -r requirements.txt --upgrade
python install.py

then restart. You need sam3.pt in ComfyUI/models/sam3/; the install.py step fetches it. Note the README's warning that the pack may try an experimental comfy-env/pixi one-click install.

Common issues & troubleshooting

It masks nothing. Your confidence_threshold is too high for that image, or the prompt doesn't match how the model "sees" the object. Drop the threshold toward 0.1, and simplify the wording - one clear noun beats a long descriptive sentence.

It masks too much / grabs the wrong stuff. Raise the threshold, or cap max_detections so you only keep the strongest hits. If there are several similar objects and you want one, add a positive_boxes prompt around the region you care about.

VRAM pressure in a bigger graph. Flip offload_model on so SAM3 leaves the GPU once the mask is out - useful when a diffusion model needs the room right after.