RecraftAI Background Remover

This node strips the background out of an image - subject stays, everything behind it goes. You feed it an image, it returns a cutout. That's the entire feature, and honestly it's the entire point of this node: background removal is a solved-enough problem that the interesting question isn't can it, it's should you pay for it.

Here's the context you need before adding this to a workflow. Background removal is the most commoditized operation in this whole ecosystem. It's free, fast, and runs locally - BiRefNet is now built into ComfyUI core, and it's the default recommendation for anything with difficult edges like hair or fur. The community's verdict on quality has moved past "can it find the subject" to "can it survive hair, fur and semi-transparency," and the local models are genuinely good at that now.

So what does Recraft's version buy you? Three things: you don't download or load any model, you don't use any VRAM, and you get Recraft's server-side matting quality - which is trained and tuned by the same team behind that leaderboard-topping image model. For a one-off cutout of a product shot, that's real convenience: wire it in, get a clean cutout, done. For a batch of 500 images, you're paying per call for something BiRefNet would do locally for free. This is the node in the pack where the "it's an API wrapper" caveat bites hardest.

The inputs and output

Required: client (from the RecraftAI Client node) and image (any IMAGE tensor - from a load node, a generator, anything). Optional: seed.

There's no prompt, no threshold, no "which object" selector. The model decides what the subject is automatically. If the image has one clear subject, this just works; if it's a busy scene with several plausible subjects, you'll get whatever it considers most salient.

The output is an image tensor with the subject cut out. Recraft returns a PNG with transparency, and the node hands it back to ComfyUI as a normal IMAGE. From there you can save it, composite it onto a new background, or feed it into other nodes. If your downstream node chokes on the alpha channel, that's a sign to check the channel count of the output before blaming the cutout.

Installing it

ComfyUI Manager: search "RecraftAI", install ComfyUI-RecraftAI, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/recraft-ai/ComfyUI-RecraftAI
cd ComfyUI-RecraftAI && pip install -r requirements.txt

Windows portable: .\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-RecraftAI\requirements.txt. Dependencies are just requests, torch, Pillow, numpy. No model downloads. You need an API key from recraft.ai/profile/api with a positive API units balance.

Where people get burned

• Paying for a free job. This is the big one. If the subject has clean edges, BiRefNet or even rembg's u2net in ComfyUI core gives you the same result at zero cost and zero latency. Reach for this node when you specifically want Recraft's edge quality on hair or fine detail and don't want to fuss with local models.
• No subject control. You can't mask which object to keep. If auto-segmentation picks the wrong thing, this node can't fix it - you'd need the Inpainter or a local segmentation route instead.
• Seed does nothing visible for most runs. It's there for API consistency, but a background removal shouldn't vary much between seeds. If it does, treat it as a signal the image is ambiguous.