LocalAI-io/sam-3d-body-dinov3-GGUF

SAM 3D Body DINOv3 — GGUF

F32 GGUF conversions of Meta's facebook/sam-3d-body-dinov3 for sam3d.cpp, a C++23/GGML CPU/Vulkan implementation. Given an image and a selected person's bounding box, the supported Body pose branch estimates a 3D body mesh, joints, pose parameters and camera translation.

This is a format conversion and component selection, not a fine-tune or low-bit quantization. GGUF names a container format: these custom SAM3D architectures require sam3d.cpp, not llama.cpp or a generic LLM runner. There is no Python, PyTorch or CUDA inference dependency in the native runtime.

Bundle

Download all three files together. The runtime validates companion identities.

The BF16 runtime mode converts selected weights/activations at load/execution; it uses these same files. There are no Q4/Q8 variants in this release. MANIFEST.json and SHA256SUMS identify the exact artifacts and conversion source commit. The model source, metadata and license files are included in the checksum manifest; no training checkpoints or reference photographs are bundled.

Download and run

Use the Hugging Face CLI after this repository has been published:

hf download LocalAI-io/sam-3d-body-dinov3-GGUF --local-dir models/sam3d-body
(cd models/sam3d-body && sha256sum -c SHA256SUMS)

Build sam3d.cpp with CMake and an installed C++23 compiler; add a Vulkan SDK for GPU inference. Its optional Go web demo accepts these paths:

./build/sam3d-demo \
  --listen 127.0.0.1:8097 --data generated/demo \
  --runner build/debug/bin/sam3d-body-infer \
  --backend CPU --module build/debug/bin/libggml-cpu.so \
  --backbone models/sam3d-body/body-dinov3-f32.gguf \
  --branch models/sam3d-body/body-pose-branch-f32.gguf \
  --mhr models/sam3d-body/mhr-lod1-f32.gguf

The demo uses the same opaque C API available to library consumers. CPU debug builds prioritize sanitizer checks; use an optimized build for performance. See sam3d.cpp's README, API guide and demo README for build, Vulkan/BF16 and resource-limit configuration. It neither downloads models nor guesses a person detector. Remote webcam access requires HTTPS and a trusted deployment.

Supported scope and limitations

• —One selected person per request, with an explicit box and camera intrinsics (the demo offers an approximate focal-length default).
• —Body pose-branch inference and MHR mesh decoding are implemented. Detailed hand-crop refinement is not implemented. These files do not reproduce the entire upstream hand-refined estimator.
• —Video uses independent frame estimates, not a learned temporal tracker. Monocular depth, camera calibration, occlusions and small/partial people can produce ambiguity or jitter. Outputs are estimates, not measurements or medical/biometric identification results.
• —Native F32 Vulkan has passed 646 recorded trained comparisons for the tested Body branch. Six strict F32 CPU geometry/projection checks remain open; BF16 has separate numerical acceptance policies and known hand-logit differences. This is not universal accuracy or full-pipeline parity across all inputs.
• —SAM 3D Objects, automatic detection, textures and animation-ready rig export are not supplied by this bundle. Demo mesh exports are static posed geometry.

Provenance and changes

• —Original weights: facebook/sam-3d-body-dinov3, revision 11aaa346c7204874a1cbafe3d39a979080b2c55a.
• —Original Body code: facebookresearch/sam-3d-body, revision b5c765a0d89d789985e186d396315e7590887b94.
• —DINOv3 code: facebookresearch/dinov3, revision 6876159a11b4df116f30f667f8c9888617df0751.
• —Geometry: independently public MHR v1.0.1, matching the upstream Body companion asset.
• —Legacy checkpoints were extracted only in an isolated trusted reference environment. Normal converters accept verified safetensors. Floating values are stored as F32, bounded topology indices as I32. Names/layout descriptors are converted for GGML; the unused DINO mask token is synthesized as zero, as required by the unmasked evaluation path. No retraining is performed.
• —The Body archive selects the pose branch, omitting hand-refinement weights; the MHR archive omits unused solver state. Source/checkpoint/config hashes are recorded in MANIFEST.json and embedded component metadata.

License and responsible use

The SAM/DINO weights are not Apache licensed. SAM material and its derivatives remain under the SAM License. DINOv3 material additionally retains the DINOv3 License. The MHR companion and topology retain MHR Apache-2.0, with applicable Momentum MIT notices. See NOTICE. The Apache grant for original sam3d.cpp code/documentation does not relicense the model or third-party adaptations.

SAM and DINO permit modification/distribution subject to their complete terms, including same-license redistribution, supplied license copies and use/legal/ trade-control restrictions. Do not use this conversion to evade upstream terms. Only process images you are authorized to use, obtain appropriate consent, and protect photographs and inferred body data. No endorsement by Meta is implied.

The original model card is authoritative for training data, intended use and research claims. Please acknowledge and cite SAM 3D Body: Robust Full-Body Human Mesh Recovery and MHR using the citations supplied by their original authors. These conversions make no independent training-data or benchmark claims.