Choose local or hosted background removal

Start with the image you want to cut out. Choose Remove backgrounds with BiRefNet when you want the background-removal model to run in your ComfyUI environment and need a transparent image assembled from the loaded image’s RGB. Choose Bria: Remove image background when you can send the image to an authenticated hosted service. Both selected recipes expose one Sweet Tea Preview Image output; neither exposes a separate mask.

Local here describes the model stage relative to the ComfyUI environment. It does not establish that the environment is on your own computer or that uploading an image to Sweet Tea involves no transfer. Bria adds a hosted provider request. If image handling is the deciding constraint, establish where your ComfyUI environment runs and whether provider transfer is acceptable before choosing. Provider retention policies are not supplied.

The retained material establishes different mechanisms and setup, but no measured winner for quality, cost or speed. Compare subject edges on an image representative of your actual task if those differences matter.

The selected BiRefNet recipe is provider version 0.1.1. Node 25, LoadBackgroundRemovalModel, resolves bg_removal_name from ComfyUI’s background_removal model category. The selected filename is birefnet.safetensors; the recipe records a download URL for that filename. The prepared requirements do not resolve its exact model artifact checksum or version, so the filename alone does not establish an exact model identity. A missing model requirement is a setup fact, not evidence that this mechanism cannot be explained.

Node 17, LoadImage, supplies its image output to node 24, RemoveBackground, through image, and to node 22, JoinImageWithAlpha, through image. Node 24 also receives node 25’s model through bg_removal_model and generates a foreground mask. That mask travels through node 23, InvertMask, before entering node 22’s alpha. JoinImageWithAlpha itself subtracts the incoming mask from one, resizing it to the image dimensions, then concatenates alpha with the image’s first three channels. The recipe’s inversion and the join’s inversion give the foreground mask the intended opacity polarity.

The prepared effect preserve_input_rgb applies to output:18 because node 22 receives the loaded image directly. This means RGB values at that stage are retained while transparency changes. It does not establish byte-for-byte preservation of the uploaded file, preservation of its original alpha, perfect segmentation or unchanged pixels after later export transformations. The generated mask is internal; the named output is an image preview.

The selected Bria recipe is provider version 0.1.6. Node 26, LoadImage, connects its image output to node 32, BriaRemoveImageBackground, through image. Node 32 returns an image to node 27, SweetTeaPreviewImage, through images; the exposed port is output:27. The retained Resource describes this as a ComfyUI partner API workflow using Bria RMBG 2.0 and requiring corresponding authenticated service access. No prompt or sampling setup is required. Authentication steps, billing terms and provider data-retention details are not supplied.

Only the loader’s image output is connected. Its separate alpha mask is not sent to Bria, so existing uploaded transparency is not preserved through this source connection. The prepared operation contains no RGB-preservation effect for Bria. This is an absence of a supported guarantee, not evidence that every RGB value changes. The selected terminal previews the returned image rather than permanently saving it through the workflow.

The prepared requirements include a resolved Resource identity for tea-time-labs/sweet-tea-nodes, but also an unresolved source-derived requirement for that pack. Exact pack version and checksum remain unknown, and the component list is incomplete. Treat successful service access and available nodes as setup needs rather than assuming a fully resolved installation.

Bria node 32 exposes moderation as the string choices "false" and "true". Selecting "true" enables the provider’s moderation mode. The retained material does not define its rejection criteria or establish a segmentation-quality effect.

The seed input controls whether the node should re-run; it does not make provider results reproducible. The selected graph uses 0, and its schema accepts integers from 0 to 2147483647. Changing it to another accepted value, such as 1, is a concrete re-run control.

There is a source conflict: the Resource prose recommends -1 for a fresh run each time, while this selected node schema has a minimum of 0. Use an accepted nonnegative value for this revision rather than relying on -1. The randomize widget entry is marked non-runtime in the retained graph; it is not an additional named provider input.

A local model problem and a hosted access problem need different remedies. Neither recipe has a retained GPU or RAM recommendation, comparable run timing, or established platform list. Both preparation records name a 1024 × 1024 standard configuration, but this is not a measured capacity requirement or an image-size limit. Use the mechanism and available setup to make the initial choice; do not infer that hosted processing eliminates every local resource need or that local processing is faster.

For a transparent cutout with retained loaded RGB, BiRefNet has the supported alpha-joining mechanism. For avoiding this local model-loading stage, Bria provides the hosted route, with service access and image transfer as the corresponding dependencies. Neither choice guarantees a correct boundary for your particular subject.