MatAnyone2 Matting (matte_ma2)

Sequence · Fast recurrent video alpha matting

Refines a binary mask seed into a soft alpha matte using a recurrent video matting network with memory propagation. The lighter, faster counterpart to matte_mama.

What you give it

• An RGB clip.
• A binary mask seed — typically from segmentation_sam3 in the same workflow.

What you get back

• A single-channel soft alpha matte clip with hair-and-edge detail.

Typical VFX use: rotoscoping a person/subject for keying, background replacement, relighting passes — wherever you want clean edges without paying the per-frame cost of a diffusion sampler.

Commercial use

Component	License	Commercial OK?
MatAnyone2 code	NTU S-Lab License 1.0 (non-commercial)	❌ No
MatAnyone2 weights	NTU S-Lab License 1.0 (non-commercial)	❌ No

Not safe for paid production work. Commercial use requires a separate agreement with NTUitive / SenseTime.

Requirements

• GPU VRAM: Substantially lower than diffusion-based matting. Expected to fit on 8–12 GB GPUs at 1080p (upstream does not publish a hard minimum).
• ComfyUI custom node: spiritform/comfy-matanyone2 (alternative: FuouM/ComfyUI-MatAnyone).
• Model weights: this plugin needs two sets of weights:
  • SAM3 checkpoint — placed manually in ComfyUI/models/checkpoints/ (default sam3.1_multiplex_fp16.safetensors). SAM3 is gated on Hugging Face: run hf auth login and accept the model terms before downloading. SAM3 is a shared dependency also used by the segmentation_sam3 plugin — see that page and the ComfyUI server setup for the one-time download and placement steps.
  • MatAnyone2 weights — matanyone2.pth (~400 MB), pulled automatically by the comfy-matanyone2 custom node. Also mirrored on Hugging Face at PeiqingYang/MatAnyone2.

Parameters

This plugin runs SAM3 segmentation and MatAnyone2 matting in a single workflow, so it exposes both sets of controls.

SAM3 Segmentation

Parameter	Meaning
Prompt	Text description of the subject to segment (e.g. person, hair, girl).
Object Indices	Comma-separated indices of the detected objects to keep.
Reference Frame	Integer index of the frame fed to SAM3 detection.
Detection Threshold	SAM3 detection confidence threshold (used by both detect and video-track stages).
Max Objects	Cap on the number of tracked objects (0 = no cap).
Detect Interval	How often (in frames) detection is re-run during tracking.
Refine Iterations	Number of mask-refinement passes on the reference frame.
Individual Masks	Emit a separate mask per detected object instead of one combined mask.

MatAnyone2

Parameter	Meaning
Frame Limit	Maximum frames loaded for MatAnyone2 (0 = no limit).
Mask Frame	Index of the frame whose SAM3 mask seeds MatAnyone2’s propagation.
Warmup Frames	Warmup frames processed before the main sequence to initialise the memory network for smoother mattes.
Erode Radius	Morphological erosion radius applied to the input mask; shrinks the boundary (0 = none).
Dilate Radius	Morphological dilation radius applied to the input mask; expands the boundary (0 = none).
Max Internal Size	Maximum resolution for internal processing (-1 = use input resolution).
Memory Frames	Frames kept in the short-term memory buffer; more = better temporal consistency, more VRAM.
Long-Term Memory	Enable long-term memory for better consistency over longer sequences.

Model

Parameter	Meaning
SAM3 Checkpoint	Checkpoint filename in ComfyUI/models/checkpoints (e.g. sam3.1_multiplex_fp16.safetensors).

Plus the standard ComfyUI base parameters.

Demos & comparisons

Side-by-side v1 vs v2 comparison. © Yang et al., S-Lab @ NTU + SenseTime, CVPR 2025/2026. Source: pq-yang.github.io/projects/MatAnyone2. Reproduced under fair-use citation; weights non-commercial (NTU S-Lab License 1.0).

Input → output

Input clip

MatAnyone 2 alpha matte

Fast recurrent video matting result. © Yang et al., S-Lab @ NTU + SenseTime, CVPR 2026 Highlight. Source: pq-yang.github.io/projects/MatAnyone2. Reproduced under fair-use citation.

• MatAnyone2 project page — pq-yang.github.io/projects/MatAnyone2 — teaser, hair/edge demonstrations, v1-vs-v2 comparisons.
• MatAnyone (v1) project page — pq-yang.github.io/projects/MatAnyone — extensive demo gallery on difficult shots.
• GitHub — pq-yang/MatAnyone2 — code and pretrained checkpoint.

Image attribution: Yang et al., S-Lab @ NTU + SenseTime, CVPR 2025/2026; reproduced for documentation purposes with citation to arXiv:2512.11782 and arXiv:2501.14677. See the credits page.

Performance

• Each frame is a single forward pass with a recurrent state — no iterative denoising. Substantially faster than diffusion-based matting.
• Interactive rates on a single mid-range GPU at 1080p in practice.
• 4K works but is slower and VRAM-hungry.

Limitations

• Seed quality matters. The binary mask from SAM3 defines the target. A poor seed (missed limbs, halo, wrong instance) propagates.
• Memory drift on long clips: the recurrent state can drift across very long shots, occlusions, or shot changes. Re-seed with a fresh mask after each cut.
• Failure modes: fast motion blur on thin structures, near-camera transparent objects, identical-color background/foreground, multi-instance ambiguity (the model tracks only the seeded subject).
• Humans-first training: non-human subjects are out-of-distribution and may degrade. Test before relying on it for animals, objects, or stylized characters.

When to use this vs MaMa

• Use MatAnyone2 for throughput, smaller VRAM budgets, and the bulk of human-subject roto/keying work.
• Use MaMa when you need maximum alpha quality on hard cases (fine hair, semi-transparent edges) and have the VRAM and time budget.

Credits

Peiqing Yang, Shangchen Zhou, Kai Hao, Qingyi Tao. MatAnyone 2: Scaling Video Matting via a Learned Quality Evaluator. CVPR 2026 (Highlight). S-Lab @ NTU + SenseTime. Paper · Project page · GitHub

Peiqing Yang, Shangchen Zhou, Jixin Zhao, Qingyi Tao, Chen Change Loy. MatAnyone: Stable Video Matting with Consistent Memory Propagation. CVPR 2025. Paper · Project page · GitHub

ComfyUI wrapper by spiritform.

Citation

@InProceedings{yang2026matanyone2,
  title     = {{MatAnyone 2}: Scaling Video Matting via a Learned Quality Evaluator},
  author    = {Yang, Peiqing and Zhou, Shangchen and Hao, Kai and Tao, Qingyi},
  booktitle = {CVPR},
  year      = {2026}
}

@InProceedings{yang2025matanyone,
  title     = {{MatAnyone}: Stable Video Matting with Consistent Memory Propagation},
  author    = {Yang, Peiqing and Zhou, Shangchen and Zhao, Jixin and Tao, Qingyi and Loy, Chen Change},
  booktitle = {CVPR},
  year      = {2025}
}