Berkeley Lab Cut Beamline Image Analysis From About a Month to Roughly 15 Minutes Using Meta's Open Vision Models

Meta has described how two of its open-weight vision models are being used in the Genesis Mission, a program run by the U.S. Department of Energy (DOE). The project, SYNAPS-I, is led by Lawrence Berkeley National Laboratory with Argonne, Brookhaven, Oak Ridge and SLAC, and targets data analysis in X-ray and neutron science.

The pipeline pairs Segment Anything Model 3, which produces pixel-level boundaries, with DINOv3, which supplies the global features used to identify each structure and place it in the sample. Running on 300 NVIDIA A100 GPUs at the National Energy Research Scientific Computing Center, the team reports beamline image analysis falling from roughly one month to about 15 minutes.

The operational detail that matters is where the weights run. The SYNAPS-I team downloads, fine-tunes and deploys the models inside its own secure computing environments, adapting models trained on natural images to scientific imagery they were never built for. An API-only model could not be used this way without moving experimental data off federal infrastructure.

Meta has also applied the same two models to assistive robotics. In a University of Pittsburgh project backed by up to $41.5 million from the Advanced Research Projects Agency for Health (ARPA-H), researchers use Segment Anything to auto-label training data for a robotic wheelchair and assistive arm across varied angles, heights, backgrounds and lighting. In both cases, the reported benefit is replacing expert manual annotation, rather than adding a new perception capability.