Meta ships Sapiens2 and Ray‑Ban update
- Meta Reality Labs published Sapiens2, a new family of human-vision models trained on 1 billion human images for pose, segmentation, normals, pointmaps, and albedo. - The models run at native 1K resolution, scale from 0.4 billion to 5 billion parameters, and add a 4K variant for higher-detail outputs. - Meta is pairing open research with consumer AI-glasses software updates as it pushes AI across labs and devices. (github.com)
Meta Reality Labs has released Sapiens2, a new family of computer-vision models built to analyze people in images at much higher detail. (arxiv.org) (github.com) Computer vision is software that turns pixels into labels and measurements. In Sapiens2, Meta is aiming at human-specific tasks such as body pose, body-part segmentation, surface normals, pointmaps, and albedo, which is a model’s estimate of a surface’s base color without lighting effects. (arxiv.org) Meta’s paper says Sapiens2 ranges from 0.4 billion to 5 billion parameters, runs at native 1K resolution, and includes hierarchical variants that support 4K images. The company says it pretrained the models on a curated dataset of 1 billion human images. (arxiv.org) (github.com) The paper reports gains over the first Sapiens release on several benchmarks: pose improved by 4 mAP, body-part segmentation by 24.3 mIoU, and normal estimation cut angular error by 45.6%. Meta also says Sapiens2 extends the system to new tasks including pointmap and albedo estimation. (arxiv.org) Those tasks are the plumbing for products that need to understand the human body precisely, including avatar systems, image editing, motion capture, and wearable cameras. Meta’s earlier Sapiens page framed the first generation as a foundation for four human-vision tasks; Sapiens2 broadens that list and pushes resolution higher. (meta.com) (arxiv.org) At the same time, Meta’s consumer glasses software has kept moving through smaller, rolling updates rather than a single splashy launch. Meta’s official AI glasses release notes show an April 9, 2026 update for Ray-Ban Meta and Oakley Meta glasses that added AI editing tools in the Meta AI mobile app. (meta.com) Those release notes describe photo animation, AI edits, and restyling effects inside the app, not a new hardware generation. A separate April 13, 2026 update for Meta Ray-Ban Display glasses added contact search, web-page summaries, adaptive display contrast, Audible support, and Be My Eyes integration. (meta.com) Meta has also been repositioning the software stack around the Meta AI app. When Meta launched that standalone app in April 2025, it said the app would become the companion app for its AI glasses and stay connected to Meta AI across devices. (about.fb.com) Taken together, the releases show Meta shipping AI in two forms at once: open research artifacts for developers and incremental features for people already wearing its glasses. The next step is not a single announcement so much as whether Meta turns Sapiens2-style perception into more visible product features. (github.com) (meta.com)