LocalAI-io/GEM-X-GGUF

GEM-X SOMA — GGUF conversion for gem-x.cpp

F32 format conversions of NVIDIA GEM-X and its observation components for gem-x.cpp, a C++23/GGML CPU/Vulkan runtime. This is an independent port, not an NVIDIA release, fine-tune, retraining, or low-bit quantization. These custom architectures require gem-x.cpp; they are not LLM GGUFs for llama.cpp. Download the bundle from LocalAI-io/GEM-X-GGUF. See download and launch instructions.

Files and compatibility

File	Contents	Bytes
gem-x-contact-f32.gguf	GEM regression model, absent-image conditioning, contact head, SOMA skeleton/identity data	177,334,400
vitpose-f32.gguf	GEM-X's DINOv3 ViT-H observation network; 77 heatmaps	3,387,861,248
yolox-f32.gguf	YOLOX-X HumanArt detector	396,106,720

SHA-256 identities of the tested artifacts are in manifest.json. Floating weights are F32; structural arrays may be integer tensors. These are format/graph conversions rather than bit-for-bit copies of the upstream archive. GEM-X's matrices are transposed/remapped to GGML layouts, ONNX initializers are selected for the inference graph, and verified rig/identity constants are included. The contact variant has 251 tensors; older 246-tensor conversions lack live absent-image conditioning and contact weights. Reconvert older files for the current default live/offline demo.

Live inference needs these three files. Offline inference additionally needs the included sam3d.cpp/ submodule's compatible Body backbone, pose branch and MHR GGUFs, with sam3d-body-infer --gem-features producing the 1024-value pose token. Those separately licensed assets are not included in this bundle. Loading all networks, workspaces and graph buffers requires more RAM/VRAM than the file sizes alone.

Model and interfaces

The temporal network has 12 blocks, a 512-wide embedding, eight attention heads and a 2048-wide feed-forward expansion. It consumes per-frame SOMA-77 image keypoints/confidences, a person box, camera intrinsics, camera angular motion, and (offline only) a compatible Body feature vector. The model selects its published observation subset internally. Native outputs include 585-value raw motion vectors, weak-perspective camera parameters, 76 local body rotations, 45 identity coefficients, 69 scale parameters, root orientation/translation, and a 77-joint skeleton. Exports are skeleton-only animated GLB, not a body mesh. See gemx.h for exact array shapes and ownership.

Live mode follows upstream's absent-image configuration: a rolling 30-frame window, two-frame warm-up and newest-frame output. It does not use SAM3D Body, offline contact refinement or offline world-trajectory reconstruction. Detection interval 1 matches upstream cadence; the demo starts at 5 and lets the user choose. Crop reuse at larger intervals changes the observation sequence.

Offline regression fixtures cover up to 120 frames. The API accepts longer sequences up to 4096 with consecutive windows; that policy is not equivalent to upstream overlapping local attention beyond 120 frames. The demo limits completed clips to 120 sampled frames.

Provenance and conversion

• GEM-X source: NVlabs/GEM-X revision 32992550dba114c62243fb55e361311972dce8f9.
• Model source: nvidia/GEM-X revision 5ccf5ca3746c3620aa4016114f069a5f6ae399cd.
• SOMA source: e0f8ff0ecfa3edbbb6058b1e0f08822ee2f84ee5; neutral rig from nvidia/SOMA-X revision 466879a.
• Detector: MMPose's YOLOX-X HumanArt a39d44ed ONNX archive.

reference/sources.json pins source URLs, byte sizes and SHA-256 identities, including checkpoint, ONNX external data and fixtures. Converters reject unexpected official artifacts. Python dependencies are conversion-only: NumPy, ONNX, gguf, and PyTorch for verified checkpoint/contact or MHR/SOMA identity extraction. Use a separate conversion environment.

python3 scripts/download_reference.py
python3 scripts/download_yolox.py
# Obtain the hash-pinned official checkpoint, SOMA rig and MHR/SOMA assets
# listed in reference/sources.json. Extract identity constants with:
python3 scripts/extract_soma_identity.py /path/to/verified/assets \
  generated/reference/soma-identity-native.npz --soma-source /path/to/SOMA
python3 scripts/convert_onnx_to_gguf.py \
  generated/reference/onnx/gem_denoiser.onnx \
  generated/reference/gem-x-contact-f32.gguf \
  --upstream-root /path/to/GEM-X \
  --soma-rig /path/to/SOMA_neutral.npz \
  --identity-data generated/reference/soma-identity-native.npz \
  --checkpoint /path/to/gem_soma.ckpt
python3 scripts/convert_vitpose_onnx_to_gguf.py \
  generated/reference/onnx/vitpose.onnx generated/reference/vitpose-f32.gguf
python3 scripts/convert_yolox_onnx_to_gguf.py \
  generated/reference/yolox-humanart.onnx generated/reference/yolox-f32.gguf
python3 distribution/verify.py --models generated/reference

The examples use an installed gguf Python package; --gguf-py can select a specific package checkout instead. Model files are ignored by source control. The manifest describes the actual tested files; a converter/toolchain change may require a newly validated manifest rather than assuming matching bytes.

Usage and evaluation

Build the native runtime as described in the README, place the files at the configured model paths and launch the demo. Runtime inference requires neither Python nor CUDA. CPU and Vulkan are supported; Vulkan tile optimizations are guarded to validated hardware, with generic fallback selection elsewhere. Strict F32 is the demo default; approximate BF16/F16 modes should not be presented as the same parity configuration.

On the pinned annotated 72-frame football clip, independent upstream/native strict-F32 comparisons yielded mean pelvis-relative joint differences of 0.88 mm offline and 0.98 mm live, with a 19.05 mm worst live joint. These measure agreement with upstream, not ground-truth pose accuracy. They are not evidence of general real-world accuracy or bit-exact upstream parity. See the parity report.

Recorded RTX 5070 Ti performance: about 48.92 ms for strict-F32 ViTPose including flip augmentation, and 15.35 fps for the simulated-camera live pipeline with detection interval 5. These differ in scope and exclude physical camera latency; see live measurements.

No training was performed for this conversion. Training/data claims belong to the upstream model card, not an independent assessment by this project. The conversion was evaluated on component fixtures and the pinned clip, not a representative demographic or clinical benchmark.

Intended use and limitations

Research, animation, motion capture and integration experiments. Monocular scale, occlusion, unusual poses, moving cameras and crop errors can degrade estimates. Crop reuse assumes the person stays near the previous detection. Hand detail and metric depth may be unreliable. This demo does not publish commands to SONIC or control a G1, and has not been validated for autonomous physical robot control, medical decisions or other safety-critical applications. Obtain permission for input footage and consider the sensitivity of movement recordings.

Licenses and attribution

The primary NVIDIA model retains the NVIDIA Open Model License, separate from the port's Apache-2.0 source license. Preserve NOTICE with a model bundle. ViTPose's DINOv3 component retains the DINOv3 license; embedded MHR data retains its MHR and applicable Momentum notices. Detector components retain their YOLOX and upstream MMPose notices. SAM3D Body is a separate dependency under Meta's terms. This card's primary license field does not replace component-specific terms.

Credit for the models and research belongs to NVIDIA, Meta and the detector and tracking authors. See source licensing/provenance.