AI Background Removal Models: SAM 3.1, Bria, BEN v2, BiRefNet Compared (2026)
Last updated August 7, 2026
Background work is now a stack: SAM 3.1 (Meta, March 2026) finds and tracks any object you name in text; matting models cut it out — Bria for licensed-data compliance, BEN v2 for hair via confidence-guided matting, BiRefNet for MIT-licensed high-res, Pixelcut for products; Depth Anything Video and LTX 2.3 Clean Plate make the composite believable. All run in invideo's agent.
Updated August 2026
AI background removal in 2026 is no longer one model with one job — it's a stack: segmentation models that find things (Meta's SAM 3.1 tracks any object you name in plain text), matting models that cut them out down to individual hairs (BEN v2, BiRefNet, Bria, Pixelcut), and scene utilities (Depth Anything Video, clean-plate tools) that make the composite believable afterwards. This guide maps the roster — what each does, how they differ on licensing and edge quality, and which to reach for — as they appear among invideo's video and image tools.
What is SAM 3.1 Video Segmentation?
SAM 3 is Meta's "unified foundation model for promptable segmentation in images and videos," and its defining upgrade is open-vocabulary text prompting: type "person, red jacket" and it exhaustively segments every instance of those concepts and tracks them through the video — no clicking frame by frame. It's an 848M-parameter model pairing a DETR-based detector with a SAM-2-style tracker, scoring 75–80% of human performance on Meta's 270K-concept SA-Co benchmark. SAM 3.1, released March 27, 2026, adds Object Multiplex — shared-memory joint multi-object tracking, faster without accuracy loss (per the official GitHub repository).
For creators, SAM 3.1 is the rotoscoping engine: isolating a subject for selective color, or generating per-object masks that downstream AI VFX steps consume. Note: SAM ships under Meta's own SAM License with gated checkpoints, not a standard open-source license.
What is Depth Anything Video for?
Not background removal — background understanding. Video Depth Anything (CVPR 2025, built on Depth Anything V2) produces temporally consistent depth maps for arbitrarily long videos, in sizes from a 28.4M-parameter Small to a 381.8M Large. A depth map turns a flat clip into a scene with geometry: depth-based blur and rack focus, relighting, parallax and 2.5D camera moves, atmosphere that sits behind the subject.
One caveat from the official release: the license splits by size — Small is Apache 2.0, Base/Large are CC-BY-NC 4.0, so commercial self-hosting of the big variants requires contacting the authors. It pairs naturally with the removers below: matte out the subject, use depth to ground whatever replaces the background.
Which background remover should you use?
Four dedicated removal/matting options, compared on what actually differentiates them:
| Model | Edge/hair quality | Video? | License & data story | Best for |
|---|---|---|---|---|
| Bria (BG Remove + Video Background Removal) | Strong (BiRefNet-based) | Yes, incl. a realtime variant | Trained exclusively on licensed data (Getty/Alamy/Envato ecosystem); commercial use via Bria agreement, with indemnification positioning | Enterprise and legal-risk-sensitive work |
| BEN v2 | Excellent — Confidence Guided Matting reprocesses low-confidence pixels | Yes, alpha-channel output | MIT open weights (94.6M params) | Hair-level matting plus video alpha |
| BiRefNet | Excellent at high res (up to 2048² variants) | Image-focused | MIT — the safest fully-open choice | Open-source self-hosting, high-res stills |
| Pixelcut Video Background Removal | Strong on product edges | Yes — frame-by-frame with transparent output | Proprietary API, free web tier | Product-photo and e-commerce convenience |
The differences that matter: Bria is the only one whose training data is fully licensed — its RMBG-2.0 model card documents 15,000+ licensed training images with bias balancing, which is why legal teams like it. BEN v2's trick is architectural: a refiner network reprocesses only the pixels the base network was unsure about — precisely where hair and fur live — at up to 4K. BiRefNet (bilateral-reference segmentation, MIT) is the open workhorse behind several commercial products, including Bria's, with variants for general use, portraits, matting, and 2048² high-res work, running ~17 FPS at 1024² on a consumer RTX 4090. Pixelcut is the pragmatist's pick: product-photo-oriented, images up to 8000×8000, plus a video mode outputting transparent-background footage.
What is LTX 2.3 Clean Plate?
As it appears in invideo's picker, LTX 2.3 Clean Plate is a scene utility from the LTX model line that generates a clean background plate — the shot without the subject — from existing footage. That's the inverse of background removal, and the missing half of many composites: once a subject is matted out, a clean plate gives you something coherent to put behind them (or erases a person or object from a shot entirely). We describe it per its in-product listing; specifics beyond that weren't independently verifiable as of August 2026, so check parameters and limits in-product. For full generative re-compositing — swapping the environment while relighting the subject — Beeble SwitchX, covered in its own guide, is the adjacent tool in the same toolbox.
When do you use which?
- Cutting a person or product out of a photo — the everyday background remover job: Bria or BiRefNet for quality, Pixelcut for e-commerce, BEN v2 when hair is the hard part.
- Removing a background from footage — the video background remover path: Bria's video model (realtime variant for live), BEN v2 for alpha exports, Pixelcut for product clips.
- Selecting specific things in video by name — SAM 3.1, a universal "find and track" engine feeding masks to everything else.
- Making the composite believable — Depth Anything Video for depth-grounded effects, LTX 2.3 Clean Plate for the empty scene behind the subject.
What people ask about background removal
What is the best AI background removal model in 2026?
There's no single winner — it splits by constraint: Bria for licensed-data compliance, BEN v2 or BiRefNet (both MIT) for open-source quality (BEN v2 leads on hair), Pixelcut for product photos, and SAM 3.1 when you need to choose which object to isolate by text prompt.
Can AI remove backgrounds from video, not just images?
Yes — as of August 2026, Bria (v3 plus a realtime variant), BEN v2 (MP4/WebM with alpha), and Pixelcut all ship dedicated video background removal, and SAM 3.1 tracks segmented objects across frames for mask-based workflows.
What's the difference between segmentation and background removal?
Segmentation (SAM 3.1) identifies and tracks which pixels belong to which object — you pick targets by text, point, or box. Matting (Bria, BEN v2, BiRefNet) produces the soft-edged alpha cutout you actually composite with. Pipelines chain them: segment to select, matte to extract.
Is there an open-source background removal model good enough for production?
Two, both MIT-licensed: BiRefNet (high-res variants to 2048², the base of several commercial products) and BEN v2 (94.6M parameters, confidence-guided matting, video alpha output). Bria's weights are CC-BY-NC — commercial use runs through a Bria agreement.
Why would I need a depth map for background work?
Because a cutout on a new background looks fake without scene geometry. Temporally consistent depth enables blur, relighting, parallax, and atmosphere that interact correctly with the subject — the difference between a sticker and a shot.
Where can you run this stack?
Every model on this page — SAM 3.1, Depth Anything Video, the four removers, and LTX 2.3 Clean Plate — is invocable from the invideo agent, and that's where the stack framing pays off: "cut me out, keep the dog, blur the street behind us" is a segmentation job, a matting job, and a depth job chained together, orchestrated from one instruction. Serious creatives get every major model in one place; the models stop being products and become verbs.
Version history: first published August 2026 — SAM 3.1 (March 2026), Video Depth Anything (CVPR 2025), Bria RMBG-2.0/VRMBG 3.0, BEN v2, BiRefNet, Pixelcut; picker-described entries flagged inline.