SAM 3D Body ONNX

ONNX Runtime exports of SAM 3D Body, organized under onnx/<variant>/.

Available variants

Variant	Path	Upstream revision	Precision
ViT-H	`onnx/vith/`	`f7f4ba4f67739e297f291639afb7c59041745367`	FP16 backbone + FP32 decoder/head graphs

DINOv3 is not published yet. Once it passes the same export, parity, and accelerator gates, it will use the sibling path onnx/dinov3/ in this same repository. ViT-H and DINOv3 need distinct ONNX files because their backbones, weights, and tensor contracts differ; they do not need separate repositories.

The ViT-H export contains 22 role-named ONNX graphs and four NumPy-native MHR artifacts. Decoder layers are separate because MHR projection and keypoint updates are interleaved between recurrent decoder steps. The exact interfaces, sizes, and SHA-256 hashes are recorded in onnx/vith/manifest.json.

Download the variant with:

hf download KitsuMate/sam-3d-body-onnx --include "onnx/vith/*" --local-dir .

This is a split-runtime export rather than one standalone ONNX graph. Consumers must follow the role schedule in the manifest and implement the documented NumPy preprocessing, camera, MHR, and hand-fusion steps. CPU golden parity is recorded in onnx/vith/parity_report.json; source identities and exported-file hashes are in onnx/vith/PROVENANCE.json.

See LICENSE for the model terms and NOTICE.md for third-party attribution.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for KitsuMate/sam-3d-body-onnx

Base model

facebook/sam-3d-body-vith

Quantized

(1)

this model