Meta's SAM 3 and DINOv3 cut a scientific imaging workflow from a month to 15 minutes

On July 21, Meta described how Segment Anything Model 3 and DINOv3 are being used at U.S. Department of Energy national laboratories to analyze scientific images close to real time. The work is one of the early SYNAPS-I projects in the Genesis Mission.

The problem is not a lack of data but an excess of it. Facilities such as Berkeley Lab's Advanced Light Source produce enormous streams of X-ray images. Meta says some detectors moved from one image every six seconds a decade ago to as many as 100,000 images per second.

The two models have different jobs

DINOv3 is a self-supervised vision model that identifies what structures represent and where they are located. SAM 3 draws their precise pixel-level boundaries. In combination, DINO supplies context and SAM supplies segmentation.

The SYNAPS-I team fine-tuned both models on scientific beamline images and ran the workflow across 300 A100 GPUs at national supercomputing facilities. The system can turn a raw scan into a semantically labeled 3D volume and return it while the experiment is still running.

About 15 minutes instead of a month

Meta's example studies how grapevines respond to drought at the cellular level. Manually labeling water-carrying xylem vessels in micro-CT scans required about a month of expert annotation for each time step; the pipeline produces a result in about 15 minutes.

The gain is not merely throughput. Instead of waiting until long after the experiment, scientists can inspect a result while measurements continue and adjust what to test next. AI becomes part of the experimental loop rather than a post-processing tool.

Why open deployment matters

Prepublication research data at national laboratories cannot simply be sent to an external cloud service. Downloadable SAM and DINO models can run on secured internal infrastructure and be adapted to the domain. In some fields, the ability to modify weights with local data matters more than a benchmark score.

These results belong to a specific imaging workflow and do not mean science now takes 15 minutes. Scientific validation, domain-expert oversight, and error analysis remain necessary. The architectural lesson is still clear: combine a component that understands context with one that extracts precise boundaries, then return the result to the real user loop.

The most useful AI product is not always the largest chat model. Sometimes it is a specialized pipeline that turns a well-defined visual task from weeks into minutes on infrastructure the user controls.