Load MVANet/BEN model by name

Remember when BEN2 launched in February 2025 with a benchmark table that supposedly made every other background remover obsolete? That hype fizzled out - the model kept working, the marketing didn't. This node is the boring, useful leftover: pick a MVANet-family model by name and it downloads the weights for you and hands you a ready-to-run model. That's the whole job, and it does it without touching your browser.

It's one of the loader nodes in Salvador E. Tropea's ComfyUI-RemoveBackground_SET pack (set-soft, the same author as ComfyUI-ImageMisc). Loaders in this pack all follow the same pattern: a dropdown of known checkpoints, an auto-download on first use, and a SET_REMBG model output you wire straight into a Remove background node. "MVANet/BEN" is really one architecture family - BEN (Background Erase Network) and BEN2 from PramaLLC are MVANet with a few activation-function tweaks, and the pack also lists the raw MVANet and Finegrain variants.

What you actually pick

The model_name dropdown has seven entries, and they're self-describing with the file size right in the name: MVANet General BEN2 F16, BEN2, BEN1 F16, BEN1, MVANet F16, MVANet, and Finegrain F16. The F16 variants are the sensible defaults - the pack's own RTX 3060 numbers put BEN2 F16 at 421ms and ~1.6GB of VRAM for a 1024px image, roughly twice the speed and half the memory of the fp32 version. Finegrain is the same architecture trained by a company on their own data; worth a shot if you're chasing edge quality, but nothing here is a BiRefNet-killer.

device defaults to AUTO, which just means "whatever ComfyUI targets" (your GPU). dtype defaults to AUTO too - same precision as the weights on disk. Drop it to float16 if you're short on VRAM; results barely change.

The outputs that matter

• model (SET_REMBG) - this is the one you connect to a processing node. Nothing else is required.
• train_w and train_h (INT) - the resolution the model was trained at (1024×1024 here). The simple Remove background node uses these automatically.
• norm_params (NORM_PARAMS) - only for the manual route, where you pre-process images yourself with the Arbitrary Normalize node from Image Misc. Ignore it for now.

Installing and a word of caution

Install the pack via ComfyUI Manager (search "Remove Background (SET)") or clone it into custom_nodes:

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

The pack's only real dependency beyond what ComfyUI already ships is seconohe - the author develops it in parallel with his nodes, so if the loader errors on first run, install a fresh copy: pip install git+https://github.com/set-soft/seconohe.git. You'll also want ComfyUI 0.3.48 or newer. On first run the node downloads BEN2 into ComfyUI/models/rembg (you can redirect that folder via the RemBG_SET key in extra_model_paths.yaml).

One known quirk worth filing away: some users report that custom nodes like this one fail to import on a cold start of ComfyUI portable, then load fine after a restart from the Manager. Annoying, but restarting fixes it.

Honest take: BEN is a good general-purpose cutter, not a magic one. If you're here because you heard BEN2 was SOTA, set your expectations - it's one solid option among many in this pack. BiRefNet is still the default for hard edges, and this node exists so you can A/B them on your own images in ten minutes.