Human4D Loader

Humans4D Loader is the setup node for the pack's image-to-SMPL path. It loads the two pieces Human4D_Img2SMPL needs - the HMR2.0 human-mesh-recovery network and a YOLO person detector - into one HUMAN4D_MODEL object. On its own it does nothing visible; it's a "load the stack" node, the way MotionDiff Loader loads the motion models. But get this right and the whole 4DHuman pipeline (video in → 3D bodies out) starts working.

The name is a little confusing at first: the model itself is HMR2.0, and "Humans4D" is the research framing the pack borrows - estimating the 3D human over time ("4D" being 3D over frames) from regular 2D footage.

How it works

On load it downloads two things: the HMR2.0 checkpoint plus its model config (from the HMR2.0 HuggingFace space), and the person detector you picked. The detector choice is yours: person_yolov8m-seg.pt (default, a segmentation-capable model tuned for people), person_yolov8s-seg.pt (smaller/faster), or the general yolov8x.pt, yolov9c.pt, yolov9e.pt if you want a beefier general object detector. Person-specific weights come from the adetailer repo; the general ones from ultralytics assets.

The inputs that matter

• detector - which YOLO weights to use. The default person_yolov8m-seg.pt is the right call for most people; step down to the -s version if detection is the bottleneck, step up if it's missing people.
• fp16 - run HMR2.0 in half precision. Off by default. If you're short on VRAM, flip it on; the quality hit is minor and mostly invisible in the rendered skeletons.

Output: HUMAN4D_MODEL, straight into Human4D_Img2SMPL.

Installation

Part of Fannovel16/ComfyUI-MotionDiff. Install via ComfyUI Manager (search "ComfyUI MotionDiff") or clone into ComfyUI/custom_nodes and install requirements, then restart. It's a big install - HMR2.0 needs timm, scikit-image, webdataset and friends - and the first load downloads the checkpoint and detector weights automatically, so give it a couple of minutes before assuming it's stuck.

Gotchas

The classic surprise is the first-load download: HMR2.0 is around a gigabyte, plus the YOLO weights, all fetched from HuggingFace on the first queue run. A half-finished download errors out; re-running resumes it.

Also, this is a fairly VRAM-hungry stack. If Human4D_Img2SMPL dies with out-of-memory, come back here and flip fp16 on and use a lighter detector - the -s person model is noticeably cheaper. And remember the loader only affects the image-estimation side of the pack; it has nothing to do with text-to-motion.