BiRefNet v1 Base – COD

COD stands for camouflaged object detection, and it's exactly as weird a specialty as it sounds: finding objects that are deliberately hard to see because they blend into their surroundings - a moth on bark, a flounder on sand, a sniper's ghillie suit in a field. This BiRefNet variant was trained on that task, which makes it genuinely different from the general-purpose cutout models.

The honest framing: for normal background removal you do not want this node. General BiRefNet assumes a clear foreground/background split and is better at it. COD is for the pathological case where the subject and background are nearly identical in color and texture, and nothing else in the family will separate them. It's the specialist's specialist.

How it works

Same removeBackground contract as the other five - runware:112@2 over the Runware REST API, image up as a data URI, cutout back as an IMAGE tensor. The camouflage-trained checkpoint does the segmenting, so the same one-input surface behaves very differently on blend-in subjects.

Inputs that matter

• image - the only input you set. Feed it your hardest, most camouflaged subject; that's what this checkpoint is for.
• outputFormat - set PNG/WEBP for alpha, same as every node here.
• outputQuality - 20–99, default 95.

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

• Using COD as your everyday remover is a mistake - it can get over-eager on low-contrast regions of normal images. It's trained to find the hidden thing, which on a regular photo can mean grabbing more than the subject.
• The community's background-removal debates never really touch COD, because nobody's cutting out camo for e-commerce. You're on your own for edge cases here - test against the general model on the specific image that matters.
• Transparency: PNG, not JPG. Yes, third time this pack has said it, because it keeps being the failure.

Keep it in the drawer until you need it. When the general BiRefNet models hand you a mush of subject-and-background, this is the node that was literally trained to find the thing that's hiding.