ComfyUI-NoBG

ComfyUI-NoBG

ComfyUI nodes for nobg, an open-source library for background removal and image matting.

Installation

Install from the Comfy Registry with comfy-cli:

comfy node install comfyui-nobg

Or install manually:

cd ComfyUI/custom_nodes
git clone https://github.com/feyninc/ComfyUI-NoBG
pip install -r ComfyUI-NoBG/requirements.txt

Requires Python ≥ 3.11 and transformers ≥ 5.5. Uses ComfyUI's existing torch and torchvision installation.

Nodes

NoBG Model Loader

Loads a model and outputs NOBG_MODEL.

• model: a Hub preset, or any folder in ComfyUI/models/nobg/ (listed as local: <name>). A folder needs a config.json.

Device selection follows ComfyUI, the checkpoint keeps its own dtype, and Hub models use the Hugging Face cache. Architecture is detected from the preset or folder name; include sam3 in a local SAM3 folder's name. SAM3 uses semantic matting.

NoBG Remove Background

Takes IMAGE (batches and video frames work) and outputs:

• cutout: RGBA IMAGE. Feed it to Save Image to get a transparent PNG.
• foreground: RGB IMAGE. These are the refined colors when refine_foreground is on.
• mask: the alpha matte as a MASK.

Options:

• refine_foreground: unmixes leftover background color from soft edges to reduce halos, using a radius of 90. Use the separate refinement node if you need a different radius.
• prompt: SAM3 only, e.g. the dog. Leave it empty to use the model's default "main subject" prompt.

Batches and video frames are processed one image at a time to limit memory use. The model moves to ComfyUI's offload device after inference.

NoBG Refine Foreground

Runs nobg's foreground color estimation on any IMAGE + MASK pair, including masks from other nodes. It outputs the refined RGB image and an RGBA cutout.

Example

Load Image ─► NoBG Remove Background ─► Save Image   (cutout → transparent PNG)
                 ▲                └──► mask → ImageCompositeMasked, etc.
NoBG Model Loader┘

Open a workflow from example_workflows/ in ComfyUI's template browser and choose your image in Load Image. For SAM3, set the prompt to the subject you want to cut out, such as the dog.

Upgrading to 0.1.2

The loader now shows only model selection. Background removal keeps only the foreground-refinement toggle and optional SAM3 text prompt; hardware overrides, batch-size tuning, download location, refinement radius, model-retention, boxes, and SAM3 aggregation/threshold widgets have been removed.

Node names and output slots are unchanged, but old widget values may not map correctly to the trimmed inputs. Re-create the loader and removal nodes in older workflows (or start from the updated examples), reconnect them, and check the refinement toggle and prompt before running. Removed API inputs no longer configure execution through ComfyUI.

Contributing

See CONTRIBUTING.md for development setup, tests, and pull-request guidelines.

License

Apache-2.0. Model weights have their own licenses: SAM3 uses Meta's SAM License.