Vision AI on the factory floor: photos, labels and welds

Two kinds of vision AI

When people say "computer vision" on a factory floor, they usually mean one of two very different things.

Which vision for which job

Simplified. Many projects use both.

	General vision models	Trained inspection systems
What they are	The image side of AI assistants like ChatGPT, Claude and Gemini	Cameras plus a model trained on images of your own parts, often with special lighting or lasers
Good at	Reading text and labels in photos, sorting and describing photos, matching a photo to a job	Finding specific defects, measuring dimensions, pass or fail at line speed
Setup	Days. No training images needed	Weeks to months. Needs labeled examples of good and bad parts
Weak at	Exact counts, exact positions, tiny or blurry text	Anything it wasn't trained on: a new part, new lighting, a new camera
Typical use	Inspection photos, travelers, packing slips, serial plates, shipping labels	Weld inspection, surface defects, gauging

Sources: Anthropic and OpenAI vision documentation; the weld studies cited below. Peekgentic summary.

What general vision models do well

A general vision model looks at a photo and answers questions about it in plain language. For a manufacturer, the useful part is usually the paperwork that lives in photos:

• Reading job travelers and labels. A photo of a job sheet or a shipping label becomes a job number, a customer and a date the system can use.
• Filing inspection photos. The model reads the traveler or serial plate in a photo and files every photo of that unit to the right job, so nobody sorts them by hand.
• Checking a shipment. A photo of a packed pallet next to the packing slip: does the label match the order?
• Describing visible damage. A first look at a freight claim or a return, for a person to confirm.

We built this for a Texas manufacturer with thousands of inspection photos that had been taken but never filed. The software reads the job sheets and serial plates in the photos, groups the photos into units, and files each unit to its job, with a person confirming anything unclear. One lesson from that project: the serial plates in real shop photos were legible only 40% to 58% of the time, so the software never trusts a plate alone. It uses the plate to confirm what the paperwork says, and anything that disagrees goes to a person.

What the AI makers say these models can't do

Both major AI makers publish the limits of their vision models, and they're worth reading before you plan a project. Anthropic's documentation for Claude says the model "might hallucinate or make mistakes when interpreting low-quality, rotated, or very small images under 200 pixels," that its location outputs "are approximate," and that it "can give approximate counts of objects in an image but might not always be precisely accurate, especially with large numbers of small objects." OpenAI's documentation lists similar limits: rotated text, small text, panoramic and fisheye images, approximate counts, and tasks "requiring precise spatial localization."

Anthropic's guidance ends with a line every manufacturer should take seriously: don't use it "for tasks requiring perfect precision or sensitive image analysis without human oversight." In plain terms: let it read and sort, and let a person decide.

Let the model read and sort. Let a person decide.

Vision for welding

Weld inspection is where vision AI gets the most attention on the factory floor, and where the gap between a research result and a working system is widest.

The research is real. A 2022 study in PLOS ONE paired a deep learning model with a laser profile sensor to measure four weld surface features on pressure vessels: weld width, reinforcement, undercut and misalignment. It reported a theoretical measurement accuracy within 0.065 mm. In industry, the vision trade publication Imaging and Machine Vision Europe reports that the automation firm DGH replaced manual checks of MIG welds on car body shells with a deep-learning vision system.

But one recent study shows the catch clearly. In 2025, researchers published WeldVGG in the journal Sensors, a model that classifies X-ray images of welds as cracking, porosity, lack of penetration, or no defect. On the public dataset it was trained on, it scored about 99.9% accuracy. On X-rays from a different public dataset it had never seen, its score (macro-F1) dropped to about 64%. Fine-tuned with just 10 labeled examples per defect type from the new images, it climbed back to 83.5% accuracy. The authors note their results "may not directly translate to unconstrained industrial environments."

Same weld model, different X-rays

WeldVGG, a 2025 weld defect classifier. Note the measures differ: accuracy on the first and last, macro-F1 on the middle.

1. Images like its training setRIAWELC test set, accuracy~99.9%
2. New X-rays it had never seenGDXray weld series, macro-F1~64%
3. After 10 labeled examples per defectGDXray, accuracy after fine-tuning83.5%

Source: López et al., WeldVGG, Sensors 25(19), 2025.

Measuring welding time: you may not need a camera

A common floor question isn't about defects at all: how much of a welder's clocked time is spent actually welding? That's called arc-on time, and the welding machine itself can often report it more exactly than a camera watching for the arc's glow. Lincoln Electric's CheckPoint and Miller's Insight are examples: both describe collecting arc-on time and weld data from the power source, viewable on a web dashboard. Cameras earn their place for what machines can't report, like parts sitting in a staging area waiting for the next step.

What that means for a fabricator:

1. A demo on someone else's welds proves little. Ask any vendor to test on your parts, your lighting and your cameras.
2. Plan to label your own examples. The jump from 64% to 83.5% came from a handful of labeled images. More good examples, from your own floor, is how these systems get reliable.
3. Keep the inspector. Use vision to flag welds for a closer look and to keep a photo record of every weld, with your certified inspector making the call.
4. Start with the paperwork around the weld. Reading the traveler, filing the weld photos to the job, and logging who inspected what is a general vision model's job, and it's ready today.

What it costs to read photos

General vision models are priced per image, and for most office uses the cost is small. Anthropic's documentation gives worked examples for a 1,000 by 1,000 pixel image:

Cost to read 1,000 photos

Model input cost only, per Anthropic's published examples. Your software, storage and review time are extra.

1. Smaller, faster modelClaude Haiku 4.5, 1 megapixel photo$1.30
2. Largest modelClaude Opus 5.5, 1 megapixel photo$5.18
3. Largest model, 4K photoClaude Opus 5.5, 3840 by 2160 pixels$19.14

Source: Anthropic, Vision documentation (platform.claude.com), image cost examples.

The model is rarely the expensive part. The work is in the software around it: getting photos off phones and shared drives, matching them to jobs, handling the unclear ones, and putting results where your team already looks.

How to start

1. Pick the photos you already take. Inspection photos, travelers, shipping labels. No new cameras.
2. Save 50 real examples, including the blurry and rotated ones, and test a vision model on them.
3. Count what it gets right with no help, and decide what a person must confirm.
4. Only then consider a trained inspection system for a specific defect, and test it on your own parts before you buy.

Our guide to automated job tracking shows how photos and labels can keep every job's status current, and photos and inspections shows what we build.