BiRefNet General 512×512 FP16

Every BiRefNet node in this pack does the same job - image in, subject cut out - and they mostly differ in which checkpoint does the cutting. Runware_birefnet_general_resolution_512x512_fp16 differs in a more interesting way: it's the fixed-resolution, half-precision variant. The model processes everything at 512×512 in FP16, which means it's faster and cheaper per image than the full-resolution siblings. The name spells out the trade honestly.

This is the one you pick for volume. Product photos, batch e-commerce cutouts, a thousand reference images for a dataset - anywhere throughput and cost matter more than pixel-perfect strands. Because background removal over the API is already cheap, the FP16 variant makes mass cutout genuinely trivial.

How it works

Identical pack plumbing: a removeBackground task (AIR runware:112@6), your image up as a data URI, cutout back as IMAGE. The fixed 512×512 working resolution is the mechanism behind the speed - the model resizes internally, segments, and you get a result that usually upscales fine.

Inputs that matter

• image - the only required input.
• outputFormat - JPG default, and for bulk work that's a real trap: JPG has no alpha. If the cutouts are going into a pipeline that composites them, set PNG once and stop thinking about it.
• outputQuality - 20–99, default 95. For bulk throughput you can drop this without most people noticing.

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

• Fixed 512×512 is the whole point and the whole limitation. Fine hair and soft edges at high source resolution are where the full-res general variant beats it. Draft with this, finalize with v1 Base.
• If your pipeline needs transparency, remember PNG - the FP16 speed advantage evaporates if you have to re-run everything because you flattened the alpha.

The one-liner: the workhorse. When the job is "cut out a lot of images, cheaply," this is the node to wire into a batch loop. When the job is "one hero image with flyaway hair," grab the full-resolution general one instead.