BiRefNet HRSOD DHU

HRSOD DHU is a mouthful that decodes to "high-resolution salient object detection, trained on the DHU dataset." In plain terms: this BiRefNet variant was trained on images where the salient object has fine detail at native resolution - texture, small features, detail that gets flattened by the 512px fixed-resolution models. If you're cutting out a subject from a large source and the smaller variants lose the fine stuff, this is the high-res specialist.

Like every node in this pack's BiRefNet row, it's the same one-input, one-output removeBackground contract - runware:112@7 running on Runware's hardware - so the choice between variants comes down entirely to what your source images look like.

How it works

A removeBackground task sent over the Runware REST API. Your image goes up as a data URI, the HRSOD-tuned checkpoint segments it at high working resolution, and the cutout comes back as a native IMAGE. Cost scales with input size, which is the honest price of high-res segmentation.

Inputs that matter

• image - the only input you set. Feed it the full-resolution source; that's where this variant's advantage lives.
• outputFormat - PNG or WEBP, or the transparency flattens in the default JPG. On a high-res cutout you almost certainly want the alpha.
• outputQuality - default 95; lower it for bulk, keep it high for a hero asset.

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

• High-res input costs more and takes longer than the fixed-resolution variants. Don't feed it a 2K image if the general model already nails it - that's the default-lane mistake.
• "High-resolution salient object detection" is about the training domain, not a promise of infinite upscale. It preserves fine detail better at source resolution; it doesn't invent resolution that isn't there.
• Test, don't assume. One run on your hardest image will tell you whether HRSOD beats v1 Base on your specific subject faster than any spec sheet.

The niche, stated plainly: big source images, fine detail that matters. When the general BiRefNet variants round off your subject's structure, HRSOD DHU is the variant that was trained to keep it.