BiRefNet Dis

BiRefNet's family tree has a bunch of weight variants, and they aren't interchangeable - each one was trained toward a different benchmark family. Runware_birefnet_dis is the variant aimed at DIS5K, the dichotomous image segmentation dataset, which is benchmark-speak for "finely structured foregrounds where the edges really matter." Thin objects, intricate silhouettes, fine detail that a casual segmentation would butcher: this is the checkpoint tuned to chase that.

For most people it won't be dramatically different from v1 Base on everyday subjects. Where it earns its slot is the hard-edge case - if you're cutting out something with delicate, well-defined structure and the general model keeps eating into it, the Dis variant is worth a run.

How it works

Same plumbing as every BiRefNet node in the pack: a removeBackground task sent to the Runware API (AIR runware:112@3), your image up as a data URI, the cutout back as an IMAGE tensor. The only thing that changes between the six variants is which checkpoint does the segmentation - the node surface is identical.

Inputs that matter

• image - the only input you'll actually set.
• outputFormat - set PNG (or WEBP) or the alpha channel dies in the default JPG. This matters twice as much on fine-edged subjects, because you want every strand's transparency intact.
• outputQuality - default 95; compression for the saved file.

Install

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

Or ComfyUI Manager → search Runware → Install → restart. API key in ComfyUI Settings → Runware API key or RUNWARE_API_KEY.

Gotchas

• "Benchmark-trained" doesn't automatically mean "better for your image." If the general model already cuts your subject cleanly, Dis is a lateral move at best. Reach for it when the general model fails on fine structure, not by default.
• The background-removal doc's honest summary applies: there's no leaderboard answer for the failures people actually care about, so test both variants on your hardest image and keep whichever wins.

Treat this as the specialist's bench option. It's one run away, costs pennies, and when your subject has edges the general model won't respect, it's exactly the tool you want in the drawer.