Researcher Jinkun Cao has released FAH (Fast and Accurate Hand Estimation) — a family of open models that estimate hand poses from a single frame or video in real-time. On a consumer-grade RTX 4070 Ti SUPER GPU, the base FAH-L model achieves a PA-MPJPE of 4.31 mm on the HOT3D benchmark at 65 frames per second with MANO mesh fitting, whereas previous-generation accurate offline estimators like HaMeR and WiLoR operate at only ~4.6–4.7 frames per second. The code is published on GitHub (noahcao/FAH) under the Apache 2.0 license, and weights are available for download on Hugging Face.

What Happened

The release is structured as a research kit: the arXiv paper 2611.07711, titled *Fast and Accurate Hand Estimation in Pixels, Interaction and World Space*, was submitted on November 5, 2026, and the code is open in a public repository. The family scales by size: the compact FAH-S runs at 101 frames per second with mesh fitting and 125 without, while the heavy FAH-H delivers the best accuracy of 4.12 mm PA-MPJPE at 27 frames per second. The model directly predicts 2D joints and dense 3D hand surface points, fits the results to parametric models MANO and/or MHR-Hand, and optionally annotates anatomical hand parts and object contact probabilities without knowing the object's geometry. A video mode with track smoothing via a causal model is provided, and a world mode additionally estimates the camera. Connecting optional modules reduces the speed of the base FAH-L: part annotation drops it to 54 frames per second, the configuration with parts and contact to 49, and video mode to 59. Export to TensorRT and Core ML is ready.

Context

Prior to this release, dense 3D hand estimation was divided into two camps: accurate offline estimators with speeds of single-digit frames per second, and fast methods that sacrificed density and 3D detail. FAH claims to shift this Pareto frontier — achieving real-time performance with offline-class accuracy simultaneously. The bet is also on the deployed output format: fitting to standard parametric hand models makes the result compatible with existing animation and robotics pipelines, while object contact probabilities point to hand-object interaction tasks, where the comparative video haco_vs_fahL on the project page indicates direct positioning against interaction benchmarks. The measurements are tied to a specific consumer-grade GPU, and FPS predictably drops as optional modules are connected — a sign of profiling discipline rather than tuning a single record number.

Why This Matters for the Industry

For the industry, FAH fills a specific engineering gap: gesture capture, real-to-sim pipelines with replay in MuJoCo, and robotic systems that require a dense 3D hand model previously had to choose between offline accuracy and real-time performance. A real-time backend with mesh fitting and contact estimation without object geometry is suitable for both gesture interfaces and edge devices, and inference costs drop by an order of magnitude compared to the offline class. The code under a permissive license allows embedding the model in research pipelines without negotiations, but for teams needing the model in commercial production, entry is blocked by the weights license: they will have to negotiate a commercial license or fine-tune the open architecture. If independent runs confirm the numbers, community builds, optimized forks, and alternative models copying this speed/accuracy compromise are expected.

Why This Matters for Users

You can try it right now: clone the repository github.com/noahcao/FAH, download the weights noahcao/FAH from Hugging Face, and run the demo with your own image, video, or live camera. Working with MANO requires separate registration at mano.is.tue.mpg.de. On RTX-class hardware, it is realistic to build a working gesture capture prototype at dozens of frames per second in one to two days, because the stated measurements were obtained on a consumer-grade GPU, not server hardware. The main practical limitation is the license: weights are distributed for research and non-commercial use only, meaning a commercial product cannot be released using these weights.

What Is Still Unknown / Limitations

Known caveats: there are no independent third-party runs yet, so the shift in the solution class relies on author measurements. PA-MPJPE is a Procrustes-aligned metric; it does not measure accuracy in world coordinates, and the only benchmark in the release materials is HOT3D, so the effectiveness of the world mode with camera estimation is not directly confirmed by the stated numbers. The weights license terms leave FAH as a fundamentally research artifact, so calling dense hand estimation a "commodity" is premature.

Sources

Author

Look at AI, editorial team