Sapiens 2 support (Meta AI, ICLR 2026) New features Sapiens 2 pose extraction — 4 model sizes: sapiens2-quick (0.4B), sapiens2-mid (0.8B), sapiens2-top (1B), sapiens2-ultra (5B). Same Goliath 308 keypoint format as v1, +4 mAP improvement. Sapiens 2 depth estimation — monocular relative depth with 4 size variants (0.4B–5B) Sapiens 2 segmentation — 29-class body-part segmentation (adds Eyeglass class vs v1's 28) Intel XPU compatible — Sapiens 2 uses standard PyTorch ops (SDPA, RMSNorm), no custom CUDA kernels. Works with Intel Arc / Xe GPUs via IPEX. Dual loading — supports SafeTensors checkpoints (via sapiens2 package) and TorchScript (.pt2) fallback Installation pip install myogait[sapiens2] # torch>=2.7 + safetensors + huggingface-hub Usage from myogait import extract # Same API as Sapiens v1, just change the model name data = extract("video.mp4", model="sapiens2-top", with_depth=True, with_seg=True) myogait run video.mp4 -m sapiens2-top --with-depth Documentation Updated README with Sapiens 2 model tables, GPU support, CLI commands Updated tutorial with Sapiens 2 installation, usage, and frame coherence scoring Updated example test_sapiens_pipeline.py with Sapiens 2 option Updated batch extraction notebook for Sapiens 2 Tests 39 new tests covering registry, model loading, heatmap conversion, depth/seg estimators, Intel XPU, constants, and extract integration All 1263 tests passing References Paper: Rawal et al., Sapiens 2: A Human Foundation Model, ICLR 2026 — arXiv:2604.21681 Code: github.com/facebookresearch/sapiens2 Models: huggingface.co/facebook/sapiens2