Summary: DOE light-source detectors now produce up to 100,000 images per second, overwhelming expert annotation capacity. The SYNAPS-I project uses fine-tuned Meta SAM 3 and DINOv3 models across 300 A100 GPUs to return reconstructed, semantically labeled 3D volumes in about 15 minutes. A grapevine demonstration cut annotation from about one month per time step to 15 minutes, though accuracy and expert-review details were not published.
Expert attention, not detector speed, is the constraint
Detectors at DOE light sources have been upgraded, and their output went from one image every six seconds to 100,000 images per second; the facilities now generate tens of petabytes each year, Meta’s AI blog reports. Photons are not what runs short here — expert attention is. Marking out the boundaries of structures within a single X-ray image can eat weeks per dataset.
SYNAPS-I exists to take that problem on. It is a multi-lab project under the Genesis Mission, and its segmentation pipeline rests on two open-weight models from Meta: SAM 3, which handles pixel-level boundaries, and DINOv3, which supplies global context — what each structure is and where it sits in the sample.
What the pipeline does in practice
Both models were fine-tuned on scientific imaging data gathered at DOE beamlines, then deployed across 300 A100 GPUs at facilities that include NERSC. What comes back is a reconstructed, semantically labeled 3D volume, handed to the scientist at the beamline while the experiment is still running — roughly 15 minutes end to end, according to the Meta blog.
Drought response in grapevines serves as the demonstration case. From micro-CT scans taken at the Advanced Light Source, the pipeline rebuilds 3D volumes of vine stems and locates xylem vessels, the tubes that carry water. Annotation that used to take about a month per time step now takes 15 minutes, Meta’s post states, and that is precisely what makes following a dynamic process at acquisition speed feasible at all.
Open weights were the deciding factor
Until a study is published, data and models stay on government infrastructure at national labs, never on external cloud services. Meta presents its open source release as the enabler: SAM and DINO can be downloaded, fine-tuned, and deployed by the SYNAPS-I team inside its own secure environment, redirecting natural-image models toward a scientific domain they were never built for.
Internalize that constraint if you build for regulated or sensitive settings. An endpoint you cannot host inside the trust boundary is out of the running, however good the benchmark numbers look. The same logic shapes decisions well beyond national labs — the tradeoffs in enforcing provider-side privacy on AI API calls are the commercial cousin of this problem.
Lessons to carry away as a builder
Three points stand out, and they are engineering observations rather than endorsements:
- Team a segmenter with a context model. Boundaries come from SAM; identity and placement come from DINOv3. On its own, neither yields a labeled volume a scientist can read.
- Domain fine-tuning is where the work sits. Both base models started out on natural images. The scientific payoff came from adapting them to beamline data.
- Latency reshapes the workflow, not merely the schedule. At fifteen minutes, analysis lands while the experiment is running — a different product from a report that arrives next month.
Limits and what remains open
Meta’s post describes the grapevine demonstration and the SYNAPS-I setup, with 60 researchers across five national labs. It publishes no accuracy numbers for the segmentation, no per-domain error rates, and no figure for 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 a labeling backlog already exists, test whether a fine-tuned open vision model can produce a first-pass mask your experts edit rather than draw. That wagers less than swapping out the workflow, and it tells you quickly whether the adaptation cost is worth paying.