SAM2Matting¶

Development line: project:sam2matting · thread sam2matting-public-release
Last event: 2026-07-15 · 1 dated since 2026-07-15 · Researched: 2026-09-08 · confidence: medium

What it is¶

SAM2Matting is an inference framework that turns foreground prompts into alpha mattes for image and video compositing.

• Matte refinement: turns masks, points, boxes, and SAM3 text prompts into fine-edge mattes.
• Video matting: combines temporal tracking with dedicated matting heads.

The released weights are 216 MB Tiny, 383 MB Base+, and 3.51 GB SAM3. Training code remains unreleased and CC BY-NC-SA 4.0 is non-commercial, so use it to evaluate inference rather than as a released training or commercial deployment stack.

Development line¶

• 2026-07-15 — SAM2Matting public project resources became available. On 2026-07-15, SAM2Matting was presented through linked official project, source-code, model-hosting, and interactive-demo resources. This marks a material public-access step for the project because readers could locate the implementation and evaluate it through the linked resources.

What changed¶

• 2026-06-22 — SAM2Matting source repository was initialized.
• 2026-06-25 — SAM2Matting arXiv v1 was submitted and its three inference checkpoints were uploaded.
• 2026-06-30 — SAM2Matting removed pip install -e . and corrected dependencies; this remains the latest upstream GitHub change.
• 2026-07-09 — SAM2Matting received a Hugging Face model card and image-segmentation tag; this changed documentation, not weights or code.
• 2026-07-15 — SAM2Matting was available as a project-code-checkpoint-demo bundle; primary histories do not establish a new upstream release on this date.

How to use this¶

1. Clone the upstream repository, create its Python 3.10 Conda environment, then run pip install -r requirements.txt. — https://github.com/FudanCVL/SAM2Matting/blob/main/README.md
2. Download an official checkpoint into checkpoints/ and match it to the SAM2 or SAM3 script family. — https://huggingface.co/FudanCVL/SAM2Matting
3. For video, provide a first-frame foreground mask, then modify the demo script's hard-coded variant, input, mask, and output paths—or wrap it—before using your own media. — https://github.com/FudanCVL/SAM2Matting/blob/main/inference_video_sam2.py
4. For interactive prompts, use interactive_sam2.py for points or boxes and interactive_sam3.py for text. — https://github.com/FudanCVL/SAM2Matting/blob/main/README.md
5. For a browser trial, use the official Space with an image and rough foreground mask; it returns an alpha matte, composite preview, and transparent PNG. — https://huggingface.co/spaces/FudanCVL/sam2matting/blob/main/app.py

Best practices¶

• Use a mask that roughly covers the foreground with white as foreground; soft grayscale masks are recommended. — https://huggingface.co/spaces/FudanCVL/sam2matting/blob/main/app.py
• Use SAM2.1-Tiny when speed matters; reserve SAM3 for cases where its prompt support justifies more VRAM and time. — https://huggingface.co/spaces/FudanCVL/sam2matting/blob/main/app.py
• For Space-hosted video, seed the first frame and keep frame count modest; that limit is specific to the ZeroGPU demo. — https://huggingface.co/spaces/FudanCVL/sam2matting/blob/main/app.py
• Do not automate the README's --save_mp4 command unchanged: current code defines it as store_false while saving is the default. — https://github.com/FudanCVL/SAM2Matting/blob/main/inference_video_sam2.py
• Keep use non-commercial or obtain permission for another use case under the stated license. — https://github.com/FudanCVL/SAM2Matting/blob/main/README.md

Superseded by this¶

• 2026-06-30 — the earlier setup instruction to add pip install -e . is obsolete; upstream removed it in favor of installing the listed requirements.

Still unknown¶

• No first-party Chinese-language page or independently reproducible Chinese practitioner report was found in the Chinese search lane; no Chinese practical recipe is asserted.
• No owned local run or independent current quality/throughput benchmark was found; paper results are not used as production expectations.
• The official Space does not expose a stable absolute timestamp for its current deployment, so its current app state is not entered as a dated development event.
• The README and current code disagree about --save_mp4; a wrapper or code fix needs testing before repeatable automation.
• Official training code, a finetuning procedure, a supported-GPU matrix, and a minimum-VRAM requirement were not found.

Sources¶

Agent brief¶

• Subject: project:sam2matting, thread sam2matting-public-release, 1 dated events 2026-07-15 → 2026-07-15.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.