Photo-to-3D has worked for one object at a time and broken down at assembly scale. A new EON AI Ventures method reconstructs a full process skid from site photographs and identifies its components from four independent sources — the engineering drawing, the photographs, the nameplates bolted to the metal, and the manufacturer’s own published catalog.

IRVINE, CA — August 3, 2026 —  EON AI Ventures today described a method that turns ordinary site photographs into a complete, correctly identified 3D model of an industrial assembly. Turning photographs into 3D already works well for a single object — one pump, one valve, one chair. It has consistently broken down when the subject is an assembly of dozens of components, and it has never carried usable identity for the parts it did build. In a measured run on one process skid, the method located 30 of 32 detected components across four photographs with zero failures, transcribed 28 equipment nameplates from the full-resolution originals, and used the valve manufacturer’s own published catalog to settle three of five disputes between the engineering drawing and the photographs. Nameplate reading, manufacturer-catalog retrieval, and the use of a published datasheet as a measuring stick inside the photograph are, to EON Reality’s knowledge, new to this field.  The technology, validation methodology, and photo-to-3D architecture are detailed in the accompanying white paper, “From Photographs to a Named Assembly – EON AI Ventures Builds Photo-to-3D at Plant Scale.” 

The wall everyone hits. Photograph a single piece of equipment and today’s reconstruction tools do a genuinely good job of it. Photograph a whole process skid or a plant area and the result falls apart: small valves, gauges, and instruments are missed entirely, some components are built as the wrong object, and nothing that does get built carries a name that the plant’s maintenance, spare-parts, or work-order systems recognize. This is well known in the industry, and it is the reason these projects stall after the first impressive demonstration.

Click on the image below to access the A Plant That Knows Its Own Name presentation.

What EON built, part one: the reconstruction. The method treats a component’s position and a component’s name as two separate jobs rather than one. Position is fully solvable from photographs: a high-resolution reference crop of each component is checked against every other view one view at a time, and agreement across views pins the object in space. Four views matter because one view gives a direction, two give a crossing, three give a point, and four give a point plus a vote — so a wrong box is identified rather than merely suspected. The same check catches two identical valves that have been wrongly merged into a single part, because a crop of one will not corroborate against a view containing only the other. Because position does not wait on naming, the geometry is corrected and delivered even where identity is still open — work that competing approaches hold back.

What EON built, part two: the identification. This is the sophisticated half, and it is the part nobody has done. Identity is established by bringing together sources that have never spoken to each other:

• The engineering drawing (P&ID) supplies the inventory of tags that are supposed to exist and, more importantly, the order in which items are connected. A drawing is a schematic, not a map, so it is used for order and never for position.
• The photographs supply existence, position, and equipment class, worked against a class library of 363 equipment classes.
• The nameplates on the equipment — read in two stages, located in reduced-resolution copies where a plate reads as text-bearing metal, then re-cropped from the full-resolution original where the characters are physically resolved. The system reads the nameplates. Nobody was doing this. A plate bolted to a part is the only statement on a skid that is not an opinion.
• Online research into the manufacturer’s own published catalog and datasheets, retrieved once per manufacturer and verified, so that what comes back is the maker’s own publication rather than a distributor’s listing.

The manufacturer’s catalog does three jobs. It is usually treated as a lookup table. Here it is a working instrument.

It settles arguments with documentary evidence. On the test skid, the drawing said gate valve and the photograph said ball valve. The system recovered the word BALON cast into the valve handle and retrieved Balon Corporation’s published range, which per the company’s own website comprises floating ball valves, trunnion ball valves, swing check valves, and needle valves — and no gate valves. The drawing was contradicted by the company that made the part. Three of the five conflicts on the skid closed that way, each with a citation attached. (Public information from balon.com. Balon Corporation is not a partner, customer, or participant in this work.)

It supplies better pictures than the field ever produced. A site photograph is taken in whatever light the plant had, from wherever the walkway allowed, with a handrail across the body of the valve. The manufacturer’s own product photography is studio-lit, shot from several angles, and unobstructed, because its entire purpose was to show the product as what it is. Once a component is identified, that imagery becomes additional build input. Reconstruction quality is no longer hostage to what a crew managed to shoot on the day. The field views are always kept and always shown first, including when they are poor; a missing or unusable view is stated as such and never quietly replaced.

It supplies the ruler. A photograph on its own has no scale — nothing in it says how large anything is. The datasheet behind the catalog entry does: it gives the component’s true dimensions in millimeters. Identify one component, look up its real size, and you have a measuring stick inside the photograph. One part of known size gives scale to everything else in the same view. That is metric scale from ordinary photographs with no survey instrument, no marker placed in the scene, and no full photogrammetric solve.

What EON built, part three: three libraries that work together. EON AI Ventures now holds three libraries that address different questions about the same part. The 3D Library describes what a part looks like in space. Smart Components describes what a part does. The Field Album records what EON has actually seen in the field — every component observed at every customer, tied to its manufacturer, its published imagery, and its real dimensions. Each entry holds the field sightings, what the part turned out to be, the manufacturer and model, the tag it carried, the datasheet dimensions, and the link to the maker’s own imagery. The next time that part appears — this site, another site, another country — it is already solved. That combination of three libraries does not exist anywhere else.

“Everyone in this business can turn a photograph of one pump into a good 3D pump. Point the same camera at a whole skid and it comes apart — half the small parts are missing, a few are built as the wrong thing, and none of them know their own name,” said Dan Lejerskar, Founder, EON Reality. “That is the wall, and it is why these projects stall. We got past it, and the way past it turned out not to be a better camera. It was reading the plate on the metal and then asking the company that made the part.”

“The Field Album is the piece I would build a decade around,” Lejerskar said. “Every job we do adds sightings, manufacturer links, and real dimensions to it. So the fiftieth site builds better than the fifth — with worse photographs — because most of what is on that skid, we have already met somewhere else. A competitor cannot catch up by working harder. They would have had to visit the sites.”

The measured results. The test installation was one complete process skid: 21 photographs, originals at 5712×4284 and 4032×3024 pixels, and one drawing sheet listing 16 tagged items. Photographic detection found 32 components. 30 of the 32 were located across the four designated views with zero failures. A deliberate control run on the smallest and hardest components returned four affirmative and five negative determinations — evidence that the check discriminates rather than agreeing with whatever is put to it. Across the 21 photographs, 94 plate-like regions were located, 29 were predicted legible, and 28 were transcribed. Three of five drawing-versus-photograph conflicts were resolved by the manufacturer’s published catalog. Before any of this reached a user, adversarial review found and closed 34 defects.

Why an engineer can act on the result. Every determination carries a confidence that comes from the evidence rule that produced it, not from a model’s opinion of its own answer. A plate physically attached to a part scores what a plate scores; two independent views agreeing score what that agreement scores. Because the same evidence yields the same number on every part, on every site, a whole installation is sortable: an engineer can see at a glance which parts of the model rest on documentary fact and which want a second look, and can rank a thousand components by how well each is known. Where two sources genuinely disagree, both claims are shown together with the single photograph that would settle it, rather than an average that neither source supports. This is what makes the output safe to integrate with the systems that dispatch technicians to real equipment.

What it means beyond oil and gas. Nothing in the method depends on the equipment being a pipe. The class library is the domain-specific part, and it is a swappable input. The same method applies to hospitals full of identical infusion pumps, ventilators, and imaging systems; to aviation maintenance with identical line-replaceable units and paperwork that must match the airframe; and to mining, utilities, shipping, and data centers, where identical motors, drives, breakers, and racks are repeated by design across hundreds of sites. The two conditions the method needs — visually identical tagged equipment, and paperwork that has drifted from what was installed — are close to universal. No measured results are claimed outside the process-equipment run described here.

Read more in the From Photographs to a Named Assembly – EON AI Ventures Builds Photo-to-3D at Plant Scale. white paper.

Learn more by tuning to our podcast.

About EON AI Ventures
EON AI Ventures is the company behind Work Intelligence — the captured, verified, and compounding knowledge of how expert work is actually done. Its Intelligence Flywheel platform (Genesis, Field IQ, Assess IQ) enables industrial enterprises to encode expert procedures into AI-guided simulations, deliver them to any worker on any device, and verify competency in the field. EON AI Ventures builds on a 25-year foundation of immersive learning technology deployed across more than 80 countries. For more information, visit www.eonaiventures.com