Prompt SEGS w/ SAM

This is the node people mean when they say "Grounded SAM" in ComfyUI: type "person" or "cat" or "red jacket", and it finds every matching thing in the image, cuts it out along its actual outline, and hands you the regions as SEGS. SimpleSyrup.PromptSEGSWithSAM does that in one node - GroundingDINO does the text-driven detection, SAM turns the boxes into precise masks - and then layers on two things the plain version doesn't give you: a negative prompt, and proper edge refinement.

Why it's the star of the pack's detection cluster: it's the node that turns natural language directly into detailer input. No YOLO weights, no class files. The whole point of text-driven detection is that "find the faces" and "find the logo" and "find the cups on the table" are the same workflow with a different string.

The pipeline, from the tooltips and the source: your positive_prompt goes to GroundingDINO, which returns boxes that score above confidence_threshold (default 0.3). Boxes can be grown or shrunk with bbox_dilation before SAM sees them. SAM generates a mask for each box. Then the negative_prompt is subtracted - regions matching text you don't want get removed from the detected area - and the surviving masks can be dilated (mask_dilation), edge-refined, and trimmed (size_threshold, keep_only / keep_by).

The two parts of this that trip people up, and what to do:

• Confidence vs. recall. confidence_threshold at 0.3 is a reasonable middle. Raise it toward 0.5 if you're getting junk detections; lower it if you're missing things. GroundingDINO is permissive, so most people tune down, not up.
• Edge refinement is on by default (refine_mask true, method GuidedFilter, detail_erode 6, detail_dilate 6). This is what keeps your masks from looking like SAM's default wobbly edges. It costs time on big images, hence mask_refinement_max_size (default 2048) bounding the working size, and execution_device (auto vs forcing CPU to dodge GPU memory pressure). If you have a ViTMatte model, the optional vitmatte_model input unlocks VITMatte as a detail_method - the highest-quality edge option, at the cost of another model.

On the output side you get segs and mask. mask is the combined prompted area as a standard ComfyUI mask - that alone is worth the node if you've ever tried to build a "mask everything that's a person" mask by hand. And because these SEGS are Impact-compatible, the regions flow straight into FaceDetailer-style detailing, the pack's regional samplers, or tiled diffusion contexts.

The negative prompt is the sleeper feature. Want "person" but not "person wearing a hat"? That's the difference between detections you have to manually prune and detections that arrive pre-pruned.

Install: the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup
cd SimpleSyrup && pip install -r requirements.txt

or ComfyUI Manager → search SimpleSyrup → Install → restart, with current ComfyUI (v3 extension API). This node needs a GroundingDINO model and a SAM model - the pack's own Load Ultralytics Model, SAM Model Loader, and a GroundingDINO loader provide them, and the downloadable-models setting can fetch known catalog entries for you. Models are the usual source of "it doesn't work": no detections → check the confidence threshold and that the prompt names something actually in the image; weird mask edges → leave refinement on and stop cranking detail_erode past what the mask needs.