MatAnyone2 SAM Loader

The interactive editor in this pack is great, but not everything wants a human clicking. If you're doing batch matting - same subject position across hundreds of clips, a fixed point in a drone shot, automated pipelines - you want the scripted path. That's what MatAnyoneSAMLoader exists for: it loads a Segment Anything model and hands it to MatAnyone2 SAM Refine, letting you build masks entirely in-node from explicit coordinates.

Think of it as the same SAM brain the editor uses, minus the point-and-click interface. Same family of checkpoints (ViT-H/B/L from Meta's Segment Anything, vendored in the pack), same results - you're just telling it where to look with numbers instead of clicks.

How it works

The loader mirrors its sibling MatAnyone Model Loader almost exactly. Three inputs: sam_model_type (vit_h default, vit_l, or vit_b), checkpoint_path (blank = auto-download to ComfyUI/models/sams/), and device (auto/cuda/mps/cpu). One output: sam_model (MATANYONE_SAM), which only plugs into SAM Refine.

Mechanically: blank path triggers a download from Meta's official dl.fbaipublicfiles.com URLs (the same sam_vit_h_4b8939.pth every SAM integration uses), the model registers from the vendored segment-anything snapshot via sam_model_registry, and it's cached in memory keyed by variant/path/device - so a second queue doesn't re-read a 2.5 GB file from disk. device on auto asks ComfyUI's model management and falls back cuda → mps → cpu.

Picking a variant

• vit_h - the default, and the obvious choice on a 12 GB+ card. Best edges, heaviest.
• vit_l - middle ground nobody really needs unless H is too big.
• vit_b - the one to reach for on a tight card. Noticeably lighter, still fine for a first-frame mask because the heavy propagation happens later in the Matte node.

Install

It's the pack, not a separate thing: ComfyUI Manager search "MatAnyone2", or:

cd ComfyUI/custom_nodes
git clone https://github.com/dreamrec/MatAnyone2_ComfyUI.git
cd MatAnyone2_ComfyUI
python -m pip install -r requirements.txt
python install.py

Restart, and SAM weights download on first use.

Gotchas

The README's troubleshooting mostly applies here by extension. The classic one: if you're already keeping SAM checkpoints somewhere, pass an explicit checkpoint_path instead of letting it re-download into models/sams/. And if your auto-download lands in the wrong place because of an older ComfyUI layout, an explicit path sidesteps the whole question.

One thing worth knowing: this node loads original SAM, not SAM 2. That's intentional - SAM2's hard mattes are exactly the thing MatAnyone exists to beat, so for a first-frame mask, original SAM's finer edges are the better starting point, and the pack's vendored snapshot guarantees the API matches.