SAM + ViTMatte Pipeline — Full Quality

A SAM mask is a selection. A matte is a fractional coverage map - the kind of alpha that survives compositing a flyaway strand of hair without a hard white fringe. Those are different problems, and the gap between them is where "good enough for a mask" dies on a green-screen comp. SAM + ViTMatte Pipeline (MEC) is the pack's end-to-end answer: SAM segments the object, iterative refinement tightens it, and a neural matting stage converts the hard mask into soft, edge-accurate alpha. If you've got a single image that needs a genuinely compositing-grade cutout, this is the node the pack built for it.

The four stages

The pipeline is four stages, all inside one node. Stage one: SAM runs from your points_json / bbox_json prompts (or existing_mask if you already have a start) and produces a coarse mask. Stage two: sam_iterations (default 2, 2–3 ideal per the tooltip) re-runs SAM with mask-derived prompts to tighten boundaries. Stage three is the matting itself - refine_method picks the backend: vitmatte (HuggingFace's neural matting, the best edge quality; needs transformers), multi_scale_guided (guided filter at three scales, best non-neural), color_aware (LAB-space), laplacian_blend, or guided_filter. auto picks the best available in that order. Stage four is cleanup: fill_holes_enabled, min_region_size (default 64) for specks, and edge_contrast (default 1.0) to sharpen boundaries under harsh lighting.

The subject presets

The genuinely thoughtful part is subject_type. Set it to custom (default) and all the widgets are honored verbatim. Set it to hair, fur, cloth, skin_face, hard_edge, or soft_glow, and it auto-tunes the trimap width and matting parameters to the subject's boundary character - hair wants a wide trimap and high detail, hard_edge (vehicles, props) wants a minimal trimap and a near-binary feel. That's the sort of preset that saves you ten minutes of knob-fiddling on every portrait.

The knobs with real teeth

The knobs with real teeth: edge_radius (default 12) is how many pixels around edges get refined - larger, softer transitions; detail_preservation (0.85) balances keeping fine detail (hair, lace) against smoothing. trimap_dilate / trimap_erode (0 = auto from edge_radius) and an optional trimap MASK let you override the auto-generated trimap for manual control. batch_mode processes every frame in the batch (off = first frame only - and for video, the SeC + MatAnyone2 pipeline is the better tool anyway).

Outputs

Outputs are rich: refined_mask (the final alpha), coarse_mask (before matting - great for before/after), edge_mask (where matting actually changed the boundary), a preview (input + alpha overlay side by side), detected_bbox, score, and info (JSON of stages, params, timings).

Install

Install: clone Code2Collapse/ComfyUI-CustomNodePacks into custom_nodes (or Manager → "CustomNodePacks"), the SAM 2.1/3 weights plus pip install git+https://github.com/facebookresearch/sam2.git, and - for the flagship vitmatte backend - pip install transformers pillow, with the model (~400 MB) auto-downloading from HuggingFace on first use. The community's standing warning applies: ViTMatte is the quality ceiling but the edge-quality gains are about the boundary, not the interior - feed it a bad coarse mask and it will politely matte your mistake.