Cam Shot Toolkit: Load SAM3 Person Detector

SAM 3D Body reconstructs one person at a time, so a photo with three people needs to be split into three crops before the mesh pass runs. This loader is how you tell the toolkit which detector does that splitting. It's optional - Process Image falls back to a built-in torchvision person detector when you don't wire one in - but it's the upgrade you reach for when the default detector starts guessing wrong, and it's the most fiddly node in the pack, entirely because Meta's SAM 3 repo is gated.

What it actually loads

The node's person_detector output feeds Process Image's optional person_detector input. The detector then finds each person, returns boxes, and Process crops around each one (with bbox_padding scaling the crop so limbs don't get chopped). SAM 3 is Meta's promptable segmentation model - the one that turns a text concept like "person" into masks for every matching instance in one pass. The gating twist: facebook/sam3 is access-gated, so repo_mode=auto (the default) tries the official repo first and silently falls back to an ungated mirror, jetjodh/sam3, if your access isn't approved. That mirror download is about 3.4 GB into ComfyUI/models/sam3_person_detector/, so the first run here is also a wait.

The inputs that matter

There are only a few you'll touch:

• implementation - transformers_sam3 (default, recommended), native_sam3 (needs Meta's separate sam3 package installed, and it pins numpy<2 which fights most ComfyUI envs), or torchvision (the lightweight Faster R-CNN fallback, no 3.4 GB download).
• repo_mode - auto (default), official, or mirror. If you've accepted Meta's terms and logged in with huggingface-cli login or HF_TOKEN, auto will use the official repo; otherwise it lands on the mirror and works anyway.
• prompt - the concept text SAM 3 searches for. Default person is right for almost everyone.
• bbox_padding - how much the crop grows around each detected box (1.0–2.0, default 1.2). Bump it if hands or feet are getting cut off in the mesh.
• mask_threshold - segmentation confidence for the SAM 3 post-processing (default 0.5). Only relevant on the transformers path.

Installation and gotchas

It's the same install as the rest of the pack - Manager search "Cam Shot Toolkit", or clone and pip install:

cd ComfyUI/custom_nodes
git clone https://github.com/EnviralDesign/sam3d-body-comfyUI-camshottoolkit
python -m pip install -r sam3d-body-comfyUI-camshottoolkit/requirements.txt

Restart, done. The two real gotchas: (1) this pack pins transformers>=5.0,<6, and the SAM 3 classes only exist in recent Transformers - if the node errors on import, your Transformers is old; (2) if you force repo_mode=official without a HF_TOKEN or a Hugging Face login that's been approved for SAM 3, you'll get a clear "could not resolve SAM3 person detector weights" error - flip to auto or mirror rather than fighting it.

When to actually use it

Honest take: for a single clean subject, skip this node entirely and let Process Image's torchvision fallback do the work - it's one less 3.4 GB download and it's plenty. Reach for it when a shot has multiple overlapping people, when the torchvision detector is misfiring, or when you want to tune the crop with prompt and bbox_padding. It's a quality-of-life switch, not a required part of the graph.