Use a background-removal mask in an image edit

Start in Generate a background-removal mask, version 0.1.2, with your source image. Its useful result is a separate mask exposed as output:20, with the role MaskPreview. It is not an edited image or a transparent cutout.

Keep the original image alongside the mask. An existing downstream flow may use the mask to choose an edit region or combine image layers, but the selected recipe stops at mask extraction and preview. No downstream connection is established by this guide.

The graph sends image output slot 0 from 17 / LoadImage to the image input of 24 / RemoveBackground. 25 / LoadBackgroundRemovalModel supplies its model output to bg_removal_model. RemoveBackground calls the model's encode_image(image) and returns the named mask output; slot 0 feeds 20 / MaskPreview through its mask input.

The model selects foreground coverage rather than returning original RGB pixels. The retained source describes a foreground mask but does not establish its numerical polarity, edge accuracy or dimensions for every input. Inspect these before handing the mask to another flow. The required model component is unresolved in the preparation record, so the default filename is not evidence that the model is installed.

Use output:20 only through a mask input and connection method supported by your existing flow. This recipe declares an output port but no downstream input port, adaptation or supported binding action, so there is no executable handoff to specify here.

For an image edit, first determine whether the receiving mask marks pixels to change or pixels to protect. To replace a background while keeping the subject, those conventions require opposite polarity. The retained pack contains no implementation of an edit-mask consumer, so its behavior must be established in that flow.

For the pinned ImageCompositeMasked implementation, mask value 0 selects destination pixels and 1 selects source pixels; intermediate values blend them. With the original image as destination and an edited image as source, protecting the subject requires zero over the subject. If your extracted foreground has high values over the subject, it needs inversion for that arrangement. Inversion is not a control in this recipe.

For a transparent cutout in an existing JoinImageWithAlpha flow, the pinned implementation uses opacity 1 - resized alpha and concatenates it with the input image's first three channels. Its alpha input therefore uses zero for opaque regions and one for transparent regions. A high-valued foreground mask needs inversion to keep the foreground opaque. This node retains input RGB at that stage, but it is absent from the selected recipe.

Segmentation answers where the foreground is. Pixel preservation depends on which image contributes pixels. The pinned masked composite combines destination × (1 - mask) with source × mask. A zero mask retains destination pixel values at that stage; a nonzero mask admits source values. A soft edge is a blend, so it does not provide the same preservation as a fully zero protected region.

In an existing flow using that implementation, retain the original as destination and use the edited image as source. The supplied preservation rule requires x = 0, y = 0, resize_source = false, an aligned mask that is zero throughout the protected region, and matching image representations. These are downstream conditions, not controls available in this mask recipe.

Check alignment against the exact images being composited. Matching width and height alone does not establish matching coordinates after a crop or repositioning. The implementation resizes the mask to source dimensions, applies placement and can resize the source; automatic resizing does not prove that a subject boundary still lines up. It also normalizes alpha representations, so representation conversion must be accounted for before claiming equality with the original.

Any preservation claim applies to that composite stage. Later resizing, color changes, encoding or other transformations need separate consideration. No composite node or preservation binding exists in the selected recipe, so extracting output:20 cannot guarantee unchanged pixels in a final image.

Separate extraction problems from downstream interpretation problems. A mask can correctly identify the foreground and still select the wrong pixels when a consumer expects the opposite polarity. Likewise, a good segmentation boundary can become displaced if the mask and images no longer share coordinates. The selected recipe supplies neither an inversion control nor a boundary-refinement control; those remedies require capabilities already supported by your downstream flow.