Industrial computer vision projects struggle because collecting and labeling rare, mission-critical images can take months and cost tens of thousands of dollars.
2
Advex uses controllable diffusion models to generate synthetic images and labels for specific edge cases selected by users.
3
In the demo, training with Advex synthetic data raised a customer's IoU from about 43% to 96% in six hours.
Summary
Vasu Sharma and Kasum Wani describe synthetic data as a way to reduce the time needed to build industrial computer vision systems. Sharma explains that common images are easy to collect, while rare cases such as defects and unusual object arrangements are often the cases that determine whether a system can go into production. Wani demonstrates Advex on a conveyor-belt use case involving boxes crossing multiple tilt trays. A user selects a region of interest, describes the desired edge case, sets the ratio of real to synthetic data, and requests a generated dataset with images and labels. In the example, the customer's IoU increased from about 43% to 96% after adding synthetic data, with the process taking six hours. The speakers also discuss feedback loops, controllable diffusion models, segmentation masks, and the decision to use off-the-shelf embedding models while focusing research on controllability.
Industrial vision systems fail on rare cases that are hard to collect
Vasu Sharma says computer vision becomes much harder outside academic benchmarks, especially in manufacturing and mission-critical applications. Real data is slow and expensive to collect, and labeling requires substantial human work. A small mistake can create a serious production problem, so teams need accuracy that ordinary datasets do not provide. The difficult examples are usually rare points outside the large cluster of common data. One such example may appear only once every few months, yet it can determine whether the system is safe to deploy. Sharma argues that teams need more examples of these cases before the model can perform reliably in production.
The usual data cycle keeps teams behind production changes
The product lifecycle described by Sharma starts with manual data collection, followed by labeling, model training, and deployment. The first deployment usually has too little data to meet the required accuracy or compliance level. As new data arrives, the model degrades because it is not updated, so the team repeats the collection, annotation, and retraining process every three to six months. This creates an ongoing loop in which engineers keep trying to catch up with production conditions. Synthetic data changes the process by providing controlled, labeled examples for the classes and edge cases that are missing from the real dataset.
Advex turns a selected edge case into a synthetic data request
Kasum Wani demonstrates Advex with a dataset of packages on conveyor-belt tilt trays. The customer's difficult case involves one box sitting across two tilt trays, which raises the question of whether one tray or both trays should move to route the box. The user selects the relevant region in the data manifold and describes the desired examples, such as single objects on multiple tilt trays. The platform also lets the user set the ratio between real and generated samples and choose how diverse the generated dataset should be. Advex then produces synthetic images and their corresponding segmentation masks.
The demo reports a large IoU increase after adding synthetic data
Wani compares the customer's model trained on real data with the result after adding Advex-generated data. The initial IoU is about 43 percent. With synthetic data generated by the platform, the reported IoU rises to 96 percent. Wani says the process took six hours and moved the system into production. He contrasts this with the five to six months the customer would have needed to collect comparable real examples. The generated samples vary in color, background, and lighting while targeting the rare arrangement that the original dataset lacked.
User feedback makes synthetic generation iterative
Advex projects generated embeddings alongside the original data so users can review the synthetic samples in the same visual space. A user can identify samples they want more of and remove samples that are unsuitable for downstream training. Wani describes this as an iterative workflow with generative agents. Each cycle with Advex takes hours in the example, while a corresponding cycle using real-world collection would take the customer several months. The purpose is to let engineers guide the generated data toward the cases that improve their model instead of accepting a fixed synthetic dataset.
Wani says Advex currently uses off-the-shelf models for embeddings while concentrating its research on controlling diffusion models. Text is a useful interface for language models, but it cannot capture all the visual details needed to control an image-generation process. Advex combines text with other inputs, including segmentation masks and customer feedback. Wani presents this combination as a way to make generated images more precise and reduce hallucinated visual content. The product's research focus is therefore on controlling generation for a particular industrial use case rather than building every model component from scratch.
Manufacturing is the first focus because rare defects are expensive to collect
Wani says Advex is focusing on manufacturing and industrial automation because these sectors have especially difficult data-collection problems. In some cases, collecting as few as 20 images can take six months. Defects and anomalies are rare by definition, so collecting enough examples for a downstream model takes a long time. Manufacturing companies also want to automate processes quickly and produce more at lower cost. Advex aims to help these factories build reliable automation by generating examples of the unusual cases that are difficult to observe in normal operations.
"The results that matter the most are the following: when the model was trained only on real data, the IoU was about 43%, but when they leveraged our synthetic data, that performance went up to 96%."Kasum Wani08:00
Who should watch
You are building a manufacturing or industrial computer vision system and rare defects are delaying deployment.
Your team spends months collecting and labeling images for unusual object arrangements or anomalies.
You want to evaluate synthetic data workflows, including user feedback and controllable image generation, without building every model component yourself.
