Open-Source remove.bg Alternative (2026, Free)

Listed models: SAM 3 and WeVisOne-VL (Tencent Apache-2.0 2B+4B perception / promptable alpha matting, added 2026-11-01). For editors rather than mask models, see Photoshop alternatives. For generation rather than cutouts, see Midjourney-class models (FLUX 3 Image Klein vs Qwen-Image-2.1-Turbo).

Sourced snapshot (2026-10-13)

Field	SAM 3
Hosts	huggingface.co/facebook/sam3 (gated); github.com/facebookresearch/sam3
HF likes / downloads	4,225 / 2,037,775
GitHub stars / forks	11,952 / 1,821
Parameters	848M (README)
License	other / SAM License (19 Nov 2025)
Hardware (README)	Python 3.12+, PyTorch 2.7+, CUDA 12.6+
Quality vs remove.bg	Not independently tested

Why SAM 3 is on a remove.bg hub

Searchers who type “open source remove.bg alternative” often want to strip a background without a per-image SaaS fee, on their own GPU. SAM 3 is the trending open model we could source on 2026-10-13: 4,225 Hugging Face likes, 848M parameters, gated facebook/sam3 card, public facebookresearch/sam3 (11,952 stars). It accepts text phrases, boxes, and clicks, and it tracks in video. That is more than “make the background transparent,” and also less than a polished consumer cutout site (no official one-click PNG product on the hosts we opened).

remove.bg is a commercial background-removal service. We use its mark only to identify the query. We did not scrape its pricing page into a comparison column.

What you actually install

The GitHub README is the install source: Python 3.12+, PyTorch 2.7+, CUDA 12.6+, then pip install -e . inside the clone. You must be accepted and authenticated on the gated Hugging Face repo before checkpoints download. Image and video predictors are separate entry points in that README (build_sam3_image_model, build_sam3_video_predictor). forkalt has not measured VRAM.

License

Hugging Face license label other. GitHub LICENSE is the SAM License, last updated 19 November 2025. It is a limited, royalty-free license with redistribution, research-acknowledgement, and trade-control terms. It is not Apache-2.0. Read the file.

What we left out

Other names that appear in this query class (rembg, BiRefNet, SAM 2, Photoshop “remove background”) are not on this list until we open their licenses and hosts in the same style. SAM 2 is a predecessor discussed in the SAM 3 README; it is not a row here.

How to choose

• Need open-vocabulary masks and tracking, and you can run CUDA: SAM 3.
• Need Tencent’s ungated Apache-2.0 perception VLM (2B + 4B) with promptable alpha matting: WeVisOne-VL. Not a hosted cutout site and not Iris-3B.
• Need a general image editor: GIMP or PhotoCraft via Photoshop alternatives (PhotoCraft vs GIMP).
• Need SVG/vector instead of a cutout: Inkscape on Adobe Illustrator alternatives.
• Need a generator, not a cutout: Iris-3B, Qwen-Image-2.1-Turbo, or MOSS-Image.

What “open” does not mean

SAM 3 weights are gated. The GitHub README is explicit: request access on facebook/sam3, then hf auth login. The code is public; the checkpoints are not anonymous. The SAM License (19 November 2025) is a limited Meta license, not Apache-2.0. If you need a consumer website that returns a transparent PNG in a browser with no GPU, this hub will not pretend SAM 3 is that website.

The architecture diagram on the SAM 3 model page is from the official README (assets/model_diagram.png). SA-CO benchmark language (“75–80% of human performance”, 270K concepts) is upstream. It is not a remove.bg accuracy score.

Related editors on this site stay on Photoshop alternatives (GIMP, PhotoCraft, PhotoCraft vs GIMP). Do not treat SAM 3 as a replacement for layers, heal, or type tools. Do not treat GIMP as a replacement for SAM 3’s open-vocabulary video tracking.

Meta’s product page also talks about Instagram Edits and Meta AI Vibes. Those are Meta product mentions, not open-source distribution channels we verified for this hub.