Cognex Deep Learning and AI: Edge AI, OneVision and VisionPro Deep Learning
VisionPro Deep Learning: Cognex's GPU AI Software, Tools and Licensing
VisionPro Deep Learning is Cognex's PC and GPU AI software, successor to ViDi: the Red, Green and Blue tools, licence tiers, versions, and when to pick it.
VisionPro Deep Learning, usually shortened to VPDL, is Cognex’s PC-based deep learning software. Cognex’s version 1.0 release notes call it “the direct successor to VisionPro ViDi”, and it carries a High Detail mode that came from SUALAB’s SuaKIT, which Cognex bought in 2019. It trains and runs models on a PC with a supported GPU, and a plant needs it, rather than the on-camera Edge AI tools, when defects are fine or numerous, when vision already runs on PCs, or when one GPU machine serves several cameras. This page covers the tools, the licensing as Cognex documents it, the version history and the choice. It belongs to Cognex deep learning and AI, in the Cognex hub.

The Red, Green and Blue tools
Cognex documentation groups the tools by colour. The colour says what question the tool answers.
| Tool | Question | Modes | Typical use |
|---|---|---|---|
| Blue Locate | Where is the feature? | Not listed in the page we read | Find parts or features to guide later tools |
| Blue Read | What does the text say? | Not listed in the page we read | Hard-to-read characters |
| Red Analyze | Where are the defects? | Focused supervised, Unsupervised anomaly, High Detail | Surface defects, scratches, contamination |
| Green Classify | Which class is this? | Focused, High Detail | Sort good, bad and defect types |
Cognex’s 1.0 notes say the former ViDi Classifier is now called Focused mode, and that High Detail is the mode that came from SuaKIT. The 2021 release was positioned for fine defects such as scratches and contaminants in medical and electronics work. VPDL stays compatible with VisionPro ViDi: Cognex notes it keeps API compatibility, can load ViDi projects and uses the same dongles, and can import SuaKIT classification projects.
Versions, dated
Cognex does not publish release dates in the documents we read, except where noted. The history as we can establish it, as of November 2026:
| Version | What changed, per Cognex release notes | Date |
|---|---|---|
| 1.0 | Successor to VisionPro ViDi; High Detail added | Not found |
| 2.0 | Announced with VisionPro 10.0 in a Cognex release; designed to integrate tightly with VisionPro 10.0 | 28 July 2021 |
| 3.0 | Neural network architecture change for High Detail; TensorRT; batch processing for Green Classify High Detail; redesigned interface | Not found |
| 3.3 | Resizing for Red Analyze High Detail; a Maximum Epochs expert parameter; mask import | Not found |
| 4.1.0 | Listed in Cognex’s release file | 27 August 2025 |
| 4.2 | Requires CUDA compute capability 7.5 or higher and a minimum GPU driver; fixes for newer NVIDIA RTX 50-series cards | Not found |
Two of these deserve a plant’s attention. The 3.0 architecture change meant older High Detail models were lost on import, so a model library has to be retrained, not just copied. And the 4.2 requirements mean GPU choice is tied to the software release: old cards may not qualify. Windows Server 2016 is not supported in 4.2, per its release notes. Read the notes for the version you will run before you order hardware.
Licensing as Cognex documents it
VisionPro and VisionPro Deep Learning require a Cognex security dongle on the PC that runs them. In client-server use only the server needs it. The licences page we read, which belongs to a 3.x documentation set, describes these tiers:
| Tier | What it allows |
|---|---|
| Standard | One GPU |
| Advanced | Up to four GPUs for training and up to two for runtime |
| ViDi base tier (older) | CPU only |
Prices are not public. Confirm the tiers for your version, and ask how a dongle is handled on a virtual machine and what the replacement process is. In our view a lost or failed dongle should be treated as a stop-the-line risk on that PC, so it belongs on the spares list.
How VisionPro Deep Learning is selected
It tends to fit when:
- defects are fine, numerous or varied enough that on-camera tools struggle, and you want the High Detail mode;
- the plant already runs VisionPro on PCs, or needs several cameras served by one GPU machine;
- you need the model-development controls of a PC tool.
On-camera Edge AI tends to fit when:
- the line is fast, the defect set is modest, and a PC at each station is not wanted;
- the team prefers one device doing both capture and judgement.
Cognex states that edge learning needs far fewer images than PC deep learning, and Cognex-linked articles advise testing both on some applications. In our view that is the honest approach: run the same labelled images through each and judge them on parts the model has not seen. The 2026 In-Sight 6900 adds a middle option, an NVIDIA Jetson controller with Advanced AI tools. Cognex has not said how those tools relate to VPDL, so we do not rank them.
A first trial, step by step
Because Cognex-linked sources advise comparing approaches, a trial should be designed to produce a fair comparison.
- Fix the optics first. Use the lighting and lens you will run in production, not a bench setup.
- Collect and label one set. Use one written defect catalogue and one reviewer for borderline images.
- Split the images into training and held-out sets before you start. Never tune on the held-out set.
- Train in the mode that fits. Red Analyze for locating defects, Green Classify for sorting, with the Focused or High Detail mode as the parts demand.
- Run the same images through the on-camera tool if the application could suit it.
- Compare on the held-out parts: escapes, false rejects, cycle time and the effort to retrain.
Illustrative scenario. A line inspects a moulded part with fine surface scratches that vary in length and angle. A team might find that an on-camera tool meets the cycle time but misses the faintest scratches, while High Detail catches them on a GPU PC at the cost of an extra computer to maintain. The decision is then a cost of escapes against a cost of ownership, and quality should make it with the data in front of them.
Where it gets hard
GPU lifecycle. The software release sets the minimum GPU. A card change, a driver update or a Windows update can break a validated system, so freeze the stack and change it under change control.
Dongles. A physical USB device in a cabinet, a virtual machine and a spare list. Plan for failure.
Model migration. The 3.0 change shows that upgrading the software can orphan trained models. Keep the training images and labels, not just the model, so you can retrain.
IT security. A PC on the OT network needs patching, backup and access control like any other. For the network side see OT networks and OT network architecture for small and medium manufacturers.
Labelling and drift. The same failure modes apply as for any AI tool; they are covered on the parent page. For the camera-based tier see Edge AI tools, for the cloud platform see OneVision, and for the rule-based PC product see VisionPro.
We have no commercial relationship with Cognex and do not resell its products. For specification and commissioning help see systems integration.
Frequently asked questions
What is VisionPro Deep Learning?
VisionPro Deep Learning, often shortened to VPDL, is Cognex's PC-based deep learning software. Cognex's version 1.0 release notes call it the direct successor to VisionPro ViDi, with a High Detail mode that came from SUALAB's SuaKIT. It trains and runs models on a PC with a supported GPU and is licensed through a Cognex security dongle.
What do the Red, Green and Blue tools do?
Per Cognex documentation, Blue Locate finds features, Blue Read reads text, Red Analyze segments images to find defects (supervised, unsupervised anomaly or High Detail modes) and Green Classify sorts images into classes (Focused or High Detail modes). The colours indicate the job: locate or read, analyse, or classify.
What is the latest VisionPro Deep Learning version?
As of November 2026 the newest version we found in Cognex documentation is 4.2, whose release notes require a GPU with CUDA compute capability 7.5 or higher and a minimum driver version. Version 4.1.0 is dated 27 August 2025 in Cognex's release file. We could not find a release date for 4.2, so check the Cognex downloads page.
How is VisionPro Deep Learning licensed?
Per Cognex documentation it needs a Cognex security dongle on the PC running it. The documented tiers are Standard, one GPU, and Advanced, up to four GPUs for training and two for runtime, plus an older ViDi base tier that is CPU only. Prices are not public. The licence page we read belongs to a 3.x documentation set, so confirm the 4.x terms.
When do I need VisionPro Deep Learning instead of Edge AI?
In our view when the defects are fine or numerous enough that on-camera tools struggle, when you already run VisionPro on PCs, or when one GPU server inspects several cameras. Cognex-linked articles advise trying both on some applications, so run the same labelled images through each and compare against real rejects.
References and further reading
- Cognex Deep Learning release notes (version 1.0) · Cognex documentationThe direct-successor statement, the High Detail mode and ViDi compatibility.
- Cognex Deep Learning release notes (version 4.2) · Cognex documentationCurrent GPU and operating system requirements and fixes.
- Cognex Deep Learning release notes (version 3.0) · Cognex documentationThe High Detail architecture change and its effect on older models.
- Cognex Deep Learning tools overview (version 1.1) · Cognex documentationThe Blue, Green and Red tools and their modes.
- Cognex Deep Learning licences (version 3.0.1) · Cognex documentationDongle, Standard and Advanced tiers, and client and server licensing.
- When to use edge learning in machine vision applications · Vision Systems DesignTrade-press guidance on choosing between edge learning and PC deep learning.