SAM3DBody RT-DETR Detector (CRT AutoDL)

Before anything in the SAM3D Body pipeline can happen, someone has to find the person in the photo. That's RT-DETR's job. RT-DETR (the Real-Time Detection Transformer from Baidu) is a plain object detector - it scans the image and returns bounding boxes - and the v4-x with an HGNet backbone is the checkpoint this pack grabs. In the SAM3D Body workflow it's the opening act: it says "there's a person here, roughly from here to here," then the SAM 3.1 checkpoint turns that box into a precise mask, and then the body model does its thing. Without the detector, the pipeline falls back to assuming one full-frame person, which is wrong as soon as your photo has a background, a second person, or a subject off-center.

CRTAutoDLSAM3BodyRTDETRDetector downloads rt_detr_v4-x-hgnet_fp16.safetensors into ComfyUI/models/diffusion_models/ and loads it through the diffusion-model path. It's the one node in the SAM3DBody family that follows the "fancy loader" pattern rather than the pure download-and-delegate one - it has real tuning inputs, because RT-DETR loads like a diffusion model and gives you the same performance switches.

The inputs that matter

• attention_method - defaults to disabled. Same story as the other CRT loader nodes: the sageattn options only do anything if you've installed sageattention yourself, and pytorch attention is the safe fallback. Leave it alone unless you're chasing speed on a known setup.
• patch_cublaslinear - off by default; toggles ComfyUI's cublas_ops fast path. Off is right for basically everyone.
• enable_fp16_accumulation - on by default, and that's fine.

The one output

MODEL - feeds the native RTDETR_detect node, which takes this model plus your image and returns bboxes. Those boxes then go into SAM3DBody_Predict's bboxes input. So this CRT node is one step removed from the visible output: it's the loader, and RTDETR_detect is the thing that actually finds people.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt

Restart ComfyUI, or use Manager (search CRT-Nodes). Needs a recent ComfyUI - RTDETR_detect lives in core comfy_extras.

Gotchas

• First run downloads ~half a gigabyte into models/diffusion_models/, console progress via CRT AutoDL, one time. Reusable with the native RT-DETR node afterward.
• It's a person detector in a 3D pack, not a general-purpose detector node. RT-DETR v4-x is COCO-classes - it detects people, but the CRT node exists to feed the SAM3D Body pipeline. For a general detection setup, ComfyUI's own RT-DETR loading is right there.
• Multi-person. RT-DETR returns multiple boxes, which is what makes SAM3DBody_Predict able to reconstruct several people in one frame - feed it and the pipeline's multi-person support actually engages.
• Updates. Red node or NaN after a pack update → right-click → Fix node (recreate).

The full chain reads like a small factory: RT-DETR finds people, SAM 3.1 masks them, MoGe-2 sets the camera, and the SAM3D Body model builds the mesh. Every piece matters, but this one is where the whole thing gets its start - and thanks to the AutoDL wrapper, the only setup you'll ever do is the initial download.