Cam Shot Toolkit: Load SAM3D Model

Every node graph needs a door, and this is the front door to the whole Cam Shot Toolkit: a node with no inputs, no settings, and one job - make sure Meta's SAM 3D Body model is present and tell the rest of the pack where it lives. You wire its model output into Cam Shot Toolkit: Process Image and never think about it again. The name is the only slightly overblown part: it doesn't load anything heavy on its own. That's deliberate, and it's why this node is so boring.

Why it exists

SAM 3D Body is Meta's single-image human reconstruction model (the same one from facebookresearch/sam-3d-body). Given one photo of a person it fits a parametric human model - body, hands, face - and returns a mesh plus pose and camera parameters. The Cam Shot Toolkit wraps it with a twist: it's tuned for calibrated pose-to-camera-shot work, meaning you can re-render that person from a new camera angle and keep the world consistent. The Load node is the "ready" check at the start of that pipeline.

Because the actual inference lives in Process Image, this loader just builds a config dict. It checks whether the weights exist locally, and if they don't it pulls them from Hugging Face before you hit the "model not found" wall two nodes later. Meta's original repo is gated, so the pack auto-downloads from an ungated mirror, jetjodh/sam-3d-body-dinov3.

What happens on first use

The download is the real event here, so budget for it:

• model.ckpt - about 2.1 GB
• assets/mhr_model.pt (the Momentum Human Rig, the parametric body) - about 700 MB

They land in ComfyUI/models/sam3dbody/. On a decent connection that's a coffee break, not a lunch break, but it will block the first run. On subsequent runs the loader just finds the files and hands you the config instantly. It also auto-detects CUDA vs CPU - CPU technically works, but you do not want to run a DINOv3-ViT-H backbone and a body decoder on CPU. The README assumes a working ComfyUI CUDA environment, and that assumption is load-bearing.

Installation

The whole pack installs the same way, and it's nothing fancy:

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

Then restart ComfyUI. Or just search "Cam Shot Toolkit" in ComfyUI Manager and let it handle the pip step. The one thing to know going in: the pack pins transformers>=5.0,<6, which is newer than what a lot of ComfyUI environments ship. If you're on an older install, the whole pack can fail to import until you bump it - so if you see transformers errors, that's the first suspect. This pack intentionally doesn't touch your existing torch/torchvision install, so no CUDA rebuild drama here.

Where people get burned

The most common failure is the first-run download dying halfway (flaky connection, disk space) and leaving a half-written model folder. The node's error message tells you exactly where the files should go and what to place there, so the fix is just: free up ~3 GB, delete the partial folder in models/sam3dbody/, and rerun. The only other gotcha is wiring - the model output is a SAM3D_MODEL socket that only connects to this pack's own Process node, so don't expect it to talk to a generic model loader. It's a key to one door. That door, however, does all the interesting work.