What changed

The remove-background guide now says the included model can drop all or part of an object a person holds (a dark microphone, phone, card or instrument), and adds a "Part of a held object is missing" row to Common problems. The CLI's default model stays u2net_human_seg. No code changes.

Why

A user reported that background removal on a person holding a dark microphone made parts of the mic transparent. u2net_human_seg is a people-only model. We measured whether a general-purpose model should replace it, with the rule to switch only if a candidate gave better edges and was no slower per frame. None was. The quality win is real, but the cheapest candidate is 2.8x slower, so the default stays and the guide now names the case.

What I measured

Candidates, with preprocessing and ONNX files exactly as rembg v2.0.85 rembg/sessions/*.py (md5 of every download matched rembg's):

Every model ran through the CLI's own createSession / process path (onnxruntime-node 1.21.1, CPU provider, graphOptimizationLevel: "all"). Model choice and per-model preprocessing were patched into a local copy only.

Ground truth. I used seven green-screen clips of people from Mixkit (Mixkit Stock Video Free License), five of them holding something: a dark handheld mic (28293), a black bass guitar (23862), black phones (28311, 28286, 24388) and a black card (9099). Clip 28287 has no object. A chroma key gave the ground-truth alpha, and each frame was composited over one of three real plates (meadow, street with bikes, dark alley). That is 45 frames per clip, 315 frames in all. Metrics: IoU of the matte at 0.5, mean absolute alpha error, the same error inside a 10 px band around the true edge, false-positive rate, and recall inside a hand-set box around each held object. No clips are committed.

Edge quality, on the 141 frames scored for all three practical models:

On all 315 frames, u2net vs isnet scored IoU 0.956 vs 0.959, edge MAE 0.109 vs 0.094, and held-object recall 0.912 vs 0.932. On the 15 frames where the full BiRefNet ran: IoU 0.978, edge MAE 0.052, against lite's 0.977 and 0.055 on the same frames.

Held-object recall by clip (all 315 frames, u2net / isnet): mic 0.982 / 0.996, card 0.780 / 0.790, bass 0.902 / 0.958, phone 0.984 / 0.983. In the contact sheets, u2net leaves see-through bands across the mic handle on some frames, drops the card entirely, and drops the guitar body. isnet keeps the objects but softens the body on the card clip, and picks up a background bike and a stand-mounted boom mic. BiRefNet-lite keeps the held objects cleanly. That last point matters for the captions skill, which relies on u2net leaving boom arms out of the matte (skills/embedded-captions/references/anti-patterns.md).

Speed, ms per frame through process(). This covers the model run plus the input/output resizes and alpha composite at 1280x720, with the first 5 frames dropped as warm-up. All models ran at the same thread settings: intraOpNumThreads: 2, interOpNumThreads: 1, one process at a time, on a shared 8-core Linux x86_64 machine under other load. The numbers are relative, not absolute.

A u2net rerun after the others gave a 632 ms median, so the box got quieter, not busier. The ratios above are, if anything, kind to the candidates. An earlier pass at onnxruntime's default thread count put isnet at 1.7x (315 frames) and BiRefNet-lite at about 14x (8-frame probe).

Verdict: BiRefNet wins on edge quality and held objects, but is 13x to 28x slower and needs about 11 GB of RAM. isnet is slightly better on edges but 2.8x slower, with new false positives. None passes "no slower", so the default stays.

What I did NOT exercise

• CoreML (the default on Apple Silicon) and CUDA. Speed ratios on those providers were not measured.
• A dedicated, quiet machine. The timings are relative only.
• The full BiRefNet on more than 15 quality frames and 10 timed frames. BiRefNet-lite was timed on 22 frames.
• Bit-exactness of mattes across thread counts. Quality frames came from both the default-thread and 2-thread runs.
• Natural footage scored against ground truth. Scored frames are green-screen composites, the chroma key is imperfect at hair and on a small spill spot, and frame-to-frame flicker was not scored. Plain (non-green) mic clips were only checked by eye.
• The end-to-end CLI video encode. ffmpeg decode and encode cost the same for every model and were left out of the timing.