Generate Mask

Produce a binary segmentation mask for any subject described in plain text. Powered by Grounded SAM (Grounding DINO + SAM) on Replicate, pinned to a fixed model version for reproducibility.

Overview

Generate Mask takes an input image plus a short text description and returns a black-and-white PNG mask isolating the described subject. White pixels mark the selected subject – the region to edit; black pixels mark the background. This white = edit polarity matches the Generate Image inpaint convention, so this node’s mask output can feed a Generate Image mask handle (or seed its Mask Painter) directly. The node also passes the original image through unchanged on a second output handle, so you can wire image + mask into a downstream inpainting node (Generate Image, Edit Image, Image to Image) without re-routing the source image.

There is no provider selection – the node uses Grounded SAM (schananas/grounded_sam) on Replicate, pinned to a specific model version so results are reproducible. The mask returned by this node can also be refined by hand in any consumer that accepts a mask input: click “Edit Mask” in the downstream config panel to open the Mask Painter pre-seeded with this output.

Configuration

Field	Type	Default	Description
Prompt	text	`""`	Text description of the subject to segment. Examples: `"the left girl's hair"`, `"the red car"`, `"the background"`. Required. Max 500 characters.
Threshold	slider	`0.3`	Detection confidence in `0.05`-`0.95` (step 0.05). Note: the current pinned model has no box-threshold input – its only mask lever is erosion/dilation – so moving this slider may not visibly change the result.

Inputs & Outputs

Inputs:

image – source image from an upstream node (Upload Image, Generate Image, Edit Image, etc.). Required.

Outputs:

image – the original input image, unchanged (passthrough). Use this to feed the same image into the downstream inpainting node alongside the mask.
mask – the generated PNG mask. White areas are the selected subject; black areas are background.

Credits

5 credits per generation (flat, no variable pricing).

Typical Workflow

The most common use is targeted inpainting – mask a subject, then re-render only that region:

Upload Image ──> Generate Mask ──┬─> Generate Image (any provider)
                                 │     prompt: "wearing a red hat"
                                 │
                                 └─> (mask handle on Generate Image)
                                       └─> mask = white(head) / black(rest)

Upload Image provides the source.
Generate Mask with prompt = "the person's head" produces a head-shaped white region on a black canvas.
Generate Image (or Edit Image / Image to Image) consumes both outputs: the image passthrough seeds the inpaint base, and the mask goes to the mask handle. See Generate Image → Inpainting & Refine.
The downstream node only re-renders pixels inside the white area; the rest of the image is preserved.

You can also use the mask output as a regular image (preview, save to storage, send to a webhook), since it is just a PNG.

Notes

The provider is Grounded SAM on Replicate – schananas/grounded_sam (Grounding DINO for text-conditioned detection + Segment Anything for the mask) – pinned to a fixed model version so the same inputs reproduce the same mask.
Segmentation quality depends on the prompt naming the subject precisely (e.g. "the left girl's hair" rather than just "hair"). It works well for distinct, clearly-described subjects, but is not perfect: an over-generic prompt or an absent subject can return an empty (fully black) mask – in that case, rewrite the prompt to be more specific.
The mask returned here can be edited by hand. Downstream nodes that accept a mask input (Generate Image, Edit Image, Image to Image with mask-supporting providers) expose an Edit Mask button that opens the Mask Painter pre-seeded with this output – useful when the auto-segmentation is almost right but needs a touch-up.
This node does not modify the image. If you want the original image to flow through, wire the image output (not just the mask) into the downstream node.