ISNet Segment

This is the node that actually does the work: point it at an image, and it separates subject from background using an ISNet model, handing you back both the cutout and the mask. It's ArtVenture's own native background-removal path - distinct from wrapping rembg or reaching for a separate BiRefNet node pack - paired with ISNetLoader (or DownloadISNetModel for a one-node fetch-and-load) for the model itself.

Background removal is, as of 2026, the most commoditized job in this whole ecosystem - free, fast, and available a dozen different ways. ISNet is a legitimate entry in that field: it's the same lineage behind isnet-general-use and isnet-anime, the improved-edge-quality options people already reach for through rembg. It won't out-cut a matting model on a wedding veil, but for an ordinary subject on a reasonably distinct background, it does the job in one node with no extra setup.

How it works

Give it an image and a loaded ISNet model, and it runs the network's segmentation over the input, thresholds the result into a foreground mask, and returns both the mask and the image with the background separated out. threshold is the confidence cutoff for that last step: lower catches more of the subject (softer edges, some risk of background bleeding in), higher is stricter (crisper cut, but can clip fine detail like flyaway hair).

The inputs and outputs that matter

Required:

• images - what you're cutting out.
• threshold (default 0.5, 0–1) - the segmentation confidence cutoff. Start at the default; nudge down if the subject is losing edges, up if background is leaking through.

Optional:

• isnet_model - a model from ISNetLoader or DownloadISNetModel. Leave it unwired and the node presumably falls back to a default, but wiring one explicitly is how you choose the general vs. anime-tuned checkpoint.
• enabled (default true) - a bypass switch. Set to false to pass the image through untouched without removing the node from your graph - handy in a template workflow where a control elsewhere decides whether background removal runs at all.
• device_mode (AUTO / Prefer GPU / CPU) - force CPU if you'd rather not spend VRAM on this; segmentation models are lightweight enough that CPU is often fine.

Outputs:

• segmented - the IMAGE with the background handled.
• mask - the MASK, useful on its own for compositing, inpainting, or feeding ImageApplyChannel to bake it back into an alpha channel.

How to install it

Part of the Art Venture pack. ComfyUI Manager: search comfyui-art-venture, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

then pip install -r comfyui-art-venture/requirements.txt and restart. No model ships with the node itself - load one via ISNetLoader or DownloadISNetModel first.

Common issues & troubleshooting

Nothing happens / image passes through unchanged. Check enabled is true. It defaults to true, but it's the first thing to check if this node seems to be doing nothing.

Edges are rough on hair, fur, or anything semi-transparent. That's the honest limit of this model class, not a setting you're missing. For genuinely difficult edges, BiRefNet (now native in ComfyUI core) or a dedicated matting model will outperform ISNet. For a clean subject on a plain background, ISNet is fine as-is.

Anime/illustration cutout looks worse than a photo cutout would. Make sure you're loading the anime-tuned checkpoint (isnetis.pth) rather than the general-purpose one - the general model is trained on photos and doesn't handle line art edges as cleanly.

Slow on CPU, or you'd rather not use GPU. device_mode controls that directly - set it to whichever you need. These models are small enough that CPU inference is usually acceptable if VRAM is the constraint.