PVL Remove Background V2 (fal.ai)

Background removal is the most commoditized operation in this whole space, and also the one where edges still get people fired up. This node is a hosted version of the current default answer: fal.ai's fal-ai/birefnet/v2 endpoint, i.e. BiRefNet, the model that became the community's go-to for hair, fur, and fine boundaries. You feed it an image, it returns the subject cut out - no model download, no models/background_removal/ folder, no VRAM.

If you already have BiRefNet running locally, this node is redundant and probably slower than your GPU. But if you're on a laptop, a shared box, or a machine where every GB of VRAM is spoken for, a per-image hosted cutout is a perfectly sane trade.

How it works

The image is base64-encoded and POSTed to fal's BiRefNet v2 endpoint with your model choice and resolution. Batches are submitted in parallel and polled together. The result comes back as a foreground image, and if you ask for it, a mask.

The model dropdown matters more than people expect. BiRefNet isn't one model:

• General Use (Light) - the default, fastest, fine for a solid subject on a contrasting background.
• General Use (Light 2K) - same idea, higher resolution input.
• General Use (Heavy) - slower, better on complex edges.
• Matting - the one for semi-transparency: veils, glass, smoke, motion blur. A plain segmentation model cannot represent a half-transparent pixel; this one predicts fractional alpha.
• Portrait - human subjects.

The knowledge base's advice applies here verbatim: for a solid subject, Light is plenty; for hair, test the Heavy/Portrait options; for anything transparent, you need Matting.

The inputs that matter

• model - pick per the above. This is the input that decides quality.
• operating_resolution - 1024x1024 or 2048x2048. Feeding a 4K image to the 1024 option silently throws away the edge detail you wanted; go 2K when the source is big.
• output_mask - off by default. Turn it on to get the MASK output; this is what you need if you're feeding the cutout into an inpaint or compositing against a different background rather than just keeping transparency.
• refine_foreground - on by default; keeps the subject's edge clean. Leave it on.
• output_format - png (default) or webp.

Outputs: foreground (IMAGE) and mask (MASK, only meaningful with output_mask on).

Installing it

Pack install, same as every node in "ComfyUI Assistant Node":

cd ComfyUI/custom_nodes
git clone https://github.com/pvlprk/comfyui-pvl-api-nodes

Restart ComfyUI, then export FAL_KEY="..." before launching. No model files - BiRefNet lives on fal's servers.

Common issues

The failure modes here are the same ones the whole background-removal genre has always had. The biggest: a perfect mask still looks pasted on without relighting - if the cutout subject doesn't match the new scene, that's a compositing problem, not this node's problem. Second, watch the operating_resolution mismatch described above; it's the most common self-inflicted wound. Third, output_mask is off by default, so if your graph expects a MASK wire and gets an error, that's why. And remember: the request (and your image) leaves your machine. For a one-off e-commerce cutout that's fine; for a 500-image dataset being processed overnight, the API bill will make you re-read the pricing page.