The bottleneck is labeling, not scanning
Upgraded detectors at DOE light sources went from one image every six seconds to 100,000 images per second, and the facilities now produce tens of petabytes annually, Meta’s AI blog reports. The scarce resource is not photons but expert attention: drawing boundaries around structures in an X-ray image can consume weeks per dataset.
That is the problem SYNAPS-I, a multi-lab project under the Genesis Mission, was set up to attack. Its segmentation pipeline runs on two open-weight Meta models: SAM 3 for pixel-level boundaries and DINOv3 for global context about what each structure is and where it sits in the sample.
What the pipeline actually does
The team fine-tuned both models on scientific imaging data collected at DOE beamlines, then deployed them across 300 A100 GPUs at facilities including NERSC. The output is a reconstructed, semantically labeled 3D volume returned to the scientist standing at the beamline while the experiment is still running — roughly 15 minutes end to end, according to the Meta blog.
The demonstration case is drought response in grapevines. Using micro-CT scans from the Advanced Light Source, the pipeline rebuilds 3D volumes of vine stems and identifies xylem vessels, the tubes that carry water. Meta’s post states that annotation which previously took about a month per time step now takes 15 minutes, which is what makes tracking a dynamic process at acquisition speed feasible at all.
Why open weights decided the deployment
National labs keep prepublication data and models on government infrastructure rather than external cloud services. Meta frames its open source release as the enabler: the SYNAPS-I team can download, fine-tune, and deploy SAM and DINO inside its own secure environment, adapting models trained on natural images to a scientific domain they were never built for.
That constraint is worth internalizing if you build for regulated or sensitive settings. A hosted endpoint you cannot run inside the trust boundary is not a candidate, no matter how good the benchmark numbers look. The same logic shapes decisions well outside national labs — the tradeoffs in enforcing provider-side privacy on AI API calls are the commercial cousin of this problem.
What to take from it as a builder
Three things stand out, and they are engineering observations rather than endorsements:
- Pair a segmenter with a context model. SAM supplies boundaries; DINOv3 supplies identity and placement. Neither alone produces a labeled volume a scientist can interpret.
- Domain fine-tuning is the work. The base models were trained on natural images. The scientific value came from adapting them on beamline data.
- Latency changes the workflow, not just the schedule. Fifteen minutes means analysis happens during the experiment, which is a different product than a report delivered next month.
Limits and open questions
Meta’s post describes the grapevine demonstration and the SYNAPS-I setup, with 60 researchers across five national labs. It does not publish accuracy numbers for the segmentation, per-domain error rates, or how much expert review still follows each 15-minute run. Those are the details that determine whether this generalizes from one sample type to the next, and they are the ones to ask for before assuming a similar pipeline fits your own imaging problem.
A practical next step: if you already have a labeling backlog, test whether a fine-tuned open vision model can produce a first-pass mask your experts edit rather than draw. That is a smaller bet than replacing the workflow, and it tells you quickly whether the adaptation cost is worth paying.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.