Most background removal tools are wrappers around the same handful of open source models. Understanding which model does what lets you pick the right tool for your specific images rather than relying on marketing claims.
Four model families cover the vast majority of background removal tools in 2026. U2-Net and its derivatives, ISNet (also called InSPyReNet or DIS), BiRefNet, and Silueta. Each was designed with a different tradeoff between speed, quality, and resource usage.
U2-Net was published in 2020 and became the default for the rembg library. It is reliable, widely deployed, and works on almost any subject. The u2netp variant is a compressed version at 4.7MB that runs on CPU in under a second.
ISNet was designed for dichotomous image segmentation, which is a fancy way of saying separating a subject from its background at high resolution. It handles edges better than U2-Net and is the model most paid tools use when they claim high quality.
BiRefNet is the newest and highest quality model. It leads on hair, fine detail, and complicated edges. The tradeoff is speed. A BiRefNet inference takes several seconds even on a decent GPU.
Silueta is a compressed version of U2-Net optimized for video frames where you need speed over absolute quality.
Human portraits with hair: BiRefNet wins clearly. ISNet is second. U2-Net is acceptable but loses flyaway strands. Silueta is not appropriate for still portraits.
Product photos with clean edges: All four models perform well. There is no meaningful difference on a boxed product against a contrasting background.
Illustrations, logos, and graphics: U2-Net and ISNet both handle these well. BiRefNet is overkill for content this simple and wastes time.
Fine detail like jewellery, lace, or mechanical components: BiRefNet wins. ISNet is close. The others struggle.
Video frames: Silueta or a compressed U2-Net variant. You cannot afford BiRefNet at thirty frames per second on consumer hardware.
u2netp runs on CPU in roughly 180 milliseconds per image. U2-Net at full size takes around 1.2 seconds on CPU. ISNet and BiRefNet are slower, in the multi-second range on CPU and under a second on a decent GPU.
This matters for the infrastructure cost of running a background removal service. A model that takes ten times longer costs ten times more in compute. For free tools, this is the difference between sustainable and not.
Remova runs a router that analyzes the image and picks the cheapest model that will produce good results. Simple images go to u2netp. Complex portraits go to BiRefNet. This is why the processing time varies between five seconds and twenty-five seconds depending on the input.
U2-Net and its derivatives are Apache 2.0 licensed. ISNet is Apache 2.0. BiRefNet is MIT. Silueta is Apache 2.0. All four can be used commercially without paying a license fee.
This is why every background removal tool uses the same underlying technology. The models are free. What you pay for is the interface, the infrastructure, and the integration.
If a tool uses U2-Net only, it will be fast but the edges on complex subjects will not be as clean as a tool using BiRefNet. If a tool claims to use a proprietary model, it is almost certainly a fine-tuned version of one of these four.
Remova exposes the model selection in the interface. If you know you are processing portraits with hair, select BiRefNet Lite. If you are processing product photos, the router will pick something fast and you will not notice the difference.
BiRefNet is the highest quality by most benchmarks. It is slower than the alternatives. For a general-purpose tool, a router that picks between them is better than a single model.
Yes. All four are available in the rembg library or as standalone ONNX models. You need Python and enough RAM. The u2netp variant runs on any machine. BiRefNet needs a GPU for reasonable speed.
You are paying for the infrastructure, the interface, and the support. Running the models yourself requires a server, maintenance, and monitoring. For most users, paying a small fee is cheaper than the time cost of running it themselves.
No. All four models assume the subject is opaque. Transparent materials like glass require a different approach called matting, which estimates per-pixel alpha values rather than a binary mask.