MatteAnything_SAMMaskFromBoxes

GroundingDINO tells you where the thing is. This node tells you what it looks like. MatteAnything_SAMMaskFromBoxes is the handoff where a bounding box becomes a precise, object-hugging mask - the moment the pipeline stops being a detector and becomes a segmenter. It's the SAM payoff, and it's the node that produces the silhouette everything else refines.

How it works

It takes the SAM_PREDICTOR from MatteAnything_InitSamPredictor and the DINO_BOXES from MatteAnything_DinoBoxes, stacks the boxes, and calls SAM's predict_torch with multimask_output=False. That last flag is the interesting one: SAM normally hedges by returning several candidate masks, but this pack asks for exactly one per box. For the two-step GroundingDINO+SAM dance, that's the right call - the box has already told SAM roughly what you want, so you don't need the menu of interpretations.

The output is a single MASK at full image resolution, one per detected box. Wire it into MatteAnything_ToTrimap and you're on the road to an actual alpha matte. Wire it straight into a compositor and you get a hard-edged cutout - which is sometimes exactly what you want, and sometimes a lie about the subject's edges.

Inputs and outputs

• sam_predictor - from MatteAnything_InitSamPredictor. Same image it was built with, or the mask won't line up with your frame.
• boxes - the DINO_BOXES from MatteAnything_DinoBoxes.
• Output: one MASK.

That's the whole node. The interesting decisions all happen in the nodes feeding it.

The empty-detection failure

Here's the failure mode that will confuse you the first time. If DinoBoxes matched nothing - vague caption, hard image - it hands this node zero boxes. SAM's predict_torch gets an empty batch and the node dies mid-graph with a stack trace that says absolutely nothing about a missing detection. The error looks like a SAM bug. It isn't; your caption just failed. Back up, reword the caption, and watch the annotated image from DinoBoxes to confirm a red box actually appeared before you debug this node.

Installing this pack

ComfyUI Manager, search Comfy_KepMatteAnything, install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/M1kep/Comfy_KepMatteAnything

then restart ComfyUI. It's the standard pack dependency trio (segment-anything, detectron2, groundingdino-py) plus the SAM weights in models/sams/:

mkdir -p ComfyUI/models/sams
# sam_vit_h_4b8939.pth into ComfyUI/models/sams/

Gotchas

• Hard edges are a feature of this step, not of matting. The mask from SAM is binary. If your subject has hair or glass, don't judge this node's output as the final result - that's literally why the pack adds ViTMatte on top.
• Multiple boxes, multiple masks. A caption that matches three objects returns three masks stacked into one MASK. If you wanted just one, fix the caption upstream, not here.

Honest take: this node is the boring, reliable bridge in a pipeline full of moving parts. It's where the phrase "Matte Anything" starts being true - GroundingDINO finds anything you can name, and SAM carves it out. Just remember the mask it hands you is the draft, not the final cut.