InstantHMR v1.1

• Kinematic fit inside the model graph: each pose is refined against the network's own 2D keypoints, on 1 or 2 crops per person (--fit 1|2, default 2). No learned parameters, fp16, runs on the GPU and on phone NPUs. PA-MPJPE on EMDB-1: medium 48.8 → 44.4 mm, nano 61.7 → 55.9 mm.
• Pose filters for video (--pose-filter one-euro|alpha-beta), tuned on 3DPW and EgoHumans: lower jitter by a third, with under half a frame of lag.
• Tracked boxes: with --detector-stride N, each person's box follows their 2D keypoints between detections.
• Faster demo: one affine warp per crop, 8-bit crops normalised inside the graph, CUDA graphs with onnxruntime-gpu; 2.8 ms per person on an RTX 4070 with TensorRT.
• New weights on Hugging Face: instanthmr-<size>-fit1.onnx and -fit2.onnx for every model. The v1.0 graphs are unchanged; --fit 0 keeps v1.0's behaviour.

InstantHMR v1.0.1

Paper preprint: https://hal.science/hal-05854387. Models: https://huggingface.co/momolesang/InstantHMR