Mask + Matting

This is the flagship. Mask Ops (MEC) is a full production matting pipeline in one node - segmenter, alpha matter, auto-quality pre-processing, despill, light wrap, video propagation, and diagnostics - with 17 outputs and enough knobs to scare anyone. The good news: the defaults are so aggressively automatic that you can ignore most of them and still get a better cutout than a hand-rolled chain.

The pipeline: a coarse mask from a segmenter, then an optional alpha matte, then polish. In this pack's terms it's MaskOpsMEC = the "Mask + Matting" node - the one-stop shop where the separate SAM/ViTMatte/RVM nodes get fused.

The parts that matter

• segmenter (default auto_best) - the coarse-mask backend: SAM 2.1, SAM 3, SAM 3.1, BiRefNet, RMBG-2.0, InSPyReNet, DIS, GroundingDINO, and more. Entries tagged [missing-deps] need an optional pip install to activate. auto_best picks what's available and best - leave it unless you know why not.
• matter (default vitmatte) - refines the alpha: none (use the segmenter mask as-is), vitmatte, rvm, bgmattingv2, matanyone, birefnet. This is where hair and fur actually get good.
• model / matter_model - pick specific weight files, or (auto) and let each backend choose.
• auto_quality (default ON) - the killer feature. It detects motion blur, low light, low contrast, speckle noise, and similar fg/bg color, then applies just-enough preprocessing (CLAHE, unsharp, NL-means, chroma stretch) before segmentation and a light guided-filter polish after. No knobs. On hard images this is the difference between a decent mask and a good one.
• auto_disambiguate (default ON) - feed positive_coords (e.g. on the face) and negative_coords (e.g. on the neck) and it scores SAM's three candidates by pos-coverage minus neg-coverage, so you get the face, not the whole person.

Then the pro stuff: despill with preserve_skin (decontaminate green/blue/whatever backing), lightwrap_strength for blending over a new background, holdout_mask (force pixels to alpha 0 - boom mics, rigs), core_mask (force pixels to alpha 1), post_refine (guided or DenseCRF), tta_flip/multiscale for quality, and robust_propagation for video - a confidence-aware re-anchor loop that re-prompts SAM when a frame's mask drifts (flow warp / DINOv2 / blend strategies).

Outputs worth wiring

The full list is long; the ones you'll actually use: alpha (the production matte), mask (coarse), preview (premultiplied), bbox/bbox_json, score (0–1 quality), info (what actually ran), and the compositing aids - despilled, lightwrap_rgba, edge_mask, inside_mask/outside_mask. problem_regions/severity/suggested_method come from the built-in failure explainer.

Install and model files

Pack install, then models:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git
pip install opencv-python>=4.7.0 scipy>=1.10.0

SAM weights go in ComfyUI/models/sams/ (SAM 3) and ComfyUI/models/sam2/ (SAM 2.1); BiRefNet/RMBG/InSPyReNet weights are auto-listed from their standard folders. auto_download (default off) can lazy-fetch missing weights from HF/torch.hub. The [missing-deps] tags tell you exactly which pip packages unlock which backends - install transformers for ViTMatte, and the extras for the video matters.

Three gotchas: (1) RMBG-2.0 is non-commercial under BRIA's license - check before shipping a product. (2) SAM 3 is not Apache like SAM/SAM2, worth knowing for commercial use. (3) With 50+ inputs, keep auto_quality on and resist touching everything - the node's own docs describe it as "no knobs to tune," and it's right.