Recraft - Remove Background

Why bother, when BiRefNet ships in ComfyUI core

That's the right first question. Locally you have a genuinely excellent stack - BiRefNet is the community default, runs around 58ms FP16 on a 4090 in ~3.5 GB, rembg bundles several BiRefNet weights, and ComfyUI ships it natively now. The quality frontier in background removal hasn't moved much since early 2025; what moved was where you can reach it.

So this node isn't about beating BiRefNet. It's for three situations: you're on a machine where 3.5 GB of VRAM is a real conversation; you want a cutout step in a pipeline that already routes other calls out and you'd rather have one billing relationship than two local installs; or you're comparing edge quality on a specific messy image and want a second opinion that isn't the same model family you already tried. Recraft is a design-focused closed lab, and this is one of its two utility endpoints that the pack exposes as a node.

The input is a URL - this is the trap

The required field here is a STRING named image, not an IMAGE socket. The tooltip says it plainly: "File URL after upload, not file content." So a normal graph looks like:

Load Image → KIE • Upload Image → (first_url) → Recraft - Remove Background → image

KIE • Upload Image takes an IMAGE and gives you first_url, and that string is what you plug into this node's image field. If you skip that step you'll be pasting a public URL by hand, which works fine too but is a strange way to spend an afternoon.

The documented constraints are the other half of the trap, because they're tighter than the upscale node's: 5 MB max (not 10), max 16 megapixels, max dimension 4096 px, minimum dimension 256 px, PNG/JPG/WEBP only. A 50-megapixel export fails on size and on pixels before anything runs.

Optional inputs are just timeout_seconds (1200 default, 30–7200) and callback_url (leave it empty for normal ComfyUI use). Outputs: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart ComfyUI, then paste your KIE key once at Settings → KIE.ai Nodes Next → Connection. It's validated before replacing a working key and stored in ComfyUI's user directory, not in the workflow - which is the behaviour you want if you share graphs on Civitai-style boards.

Notes from the background-removal world that apply here

Do not use an API cutout to build a LoRA training set. The KB's guidance on this is blunt and it doesn't care whether the mask came from BiRefNet or a cloud service: clipping a subject against transparency teaches the trainer that your subject lives in a void, and the background edges bleed.

If the result comes back and your graph's alpha looks suspicious, check the url output rather than the tensor first. The remote result is what Recraft produced; what reaches the IMAGE socket is whatever the pack downloaded and decoded, and comparing the two takes ten seconds and tells you which side you're debugging.

And the honest trade: on hair, fur and veils, the community has never settled BiRefNet vs InSPyReNet because there's no shared benchmark, so treat any single result as one opinion. If a cutout looks bad on wispy edges, that's the known hard case for the entire category - not a sign you configured this node wrong.