Load IS-Net model by name

IS-Net is what you reach for when you need genuinely good edges but don't want to pay the BiRefNet tax. It's the direct evolution of U-2-Net - same nested-UNet idea, no Swin backbone - and it's the model behind the isnet-general-use weight that carried rembg for years. The pack's benchmark numbers are almost embarrassing for the competition: 196ms and 776MB of VRAM at 1024×1024, against BiRefNet's 516ms and 1.6GB. For batch jobs, that difference is the whole point.

The loader gives you five entries:

• IS-Net Base (isnet) (169 MiB) - the plain checkpoint.
• IS-Net DIS5K (isnet-general-use) (169 MiB) - trained on the DIS5K benchmark set. This is the rembg favorite and the one most people actually mean by "isnet." Default pick.
• IS-Net CarveSet (isnet-97-carveset) - the carve-set trained variant, aimed at finer object boundaries.
• IS-Net Anime (ISNet_anime-seg) (195 MiB, 640×640) - the anime/illustration specialist. If you're cutting out flat-colour artwork, this is the one to try; the general models are tuned for photography.
• IS-Net BRIA v1.4 - the BRIA RMBG 1.4 checkpoint. Proprietary dataset, not for commercial use - the loader prints a console warning if you select it, and it's worth respecting.

Inputs and outputs

Same loader pattern as the rest of the pack: model_name, device (AUTO = your GPU), dtype (AUTO = match weights; these ship fp32, so forcing float16 is your free VRAM discount). Outputs are model (SET_REMBG) for the processing nodes, train_w/train_h (1024×1024 for the general ones, 640×640 for anime), and norm_params for manual pre-processing.

Install

Nothing special - same pack, same steps:

cd ComfyUI/custom_nodes
git clone https://github.com/set-soft/ComfyUI-RemoveBackground_SET
pip install -r ComfyUI-RemoveBackground_SET/requirements.txt

ComfyUI 0.3.48+, seconohe >= 1.0.6 (fresh pip install git+https://github.com/set-soft/seconohe.git if it errors). Weights auto-download into ComfyUI/models/rembg on first use.

Honest verdict: IS-Net is the pack's best "fast and good enough" loader, and it doubles as the only real path to good illustration cutouts thanks to the anime weight. It won't beat BiRefNet HR on a 4K portrait with frizzy hair - nothing in this budget class does. But for product shots, batch catalog work, or a 4090 user who doesn't feel like waiting, this is the one to standardize on. U-2-Net is faster still, but IS-Net's edge quality is a big enough jump that the extra 50ms is money well spent.