WACV 2026: The Practitioner’s Guide to Synthetic Data and AI Forensics

WACV 2026’s biggest trend: synthetic reality and the "cold start" data problem

At the Winter Conference on Applications of Computer Vision 2026, the discussion is shifting from data collection to automated data synthesis. With the cost of high-fidelity manual labeling scaling linearly and the demand for diverse edge cases growing exponentially, synthetic data is no longer optional for training robust production models.

This year’s WACV program is anchored by the SAFE 2026 (Synthetic & Adversarial ForEnsics) workshop. As we integrate generative models into our training loops, engineers face a new set of "forensic" challenges:

• Artifact Detection: Identifying systematic biases or visual artifacts introduced by the rendering engine or GAN that the model might "cheat" on during training.
• Adversarial Defense: Ensuring that synthetic training doesn't leave the model vulnerable to adversarial perturbations in real-world deployments.
  
  

The WACV 2026 SAFE workshop is scheduled for March 6, 2026, the first day of workshops, and has been organised by Josué Martínez-Martínez, Artificial Intelligence Researcher at MIT.

However, despite the promise of infinite data, CV teams consistently report that "infinite data ≠ infinite accuracy." To move a model from a simulator to a warehouse, engineers must navigate three primary hurdles:

• Covariate shift: The statistical mismatch between the synthetic "perfect" source domain and the noisy, imperfect target domain (e.g., sensor noise, motion blur, and varied lighting).
• Lack of semantic diversity: Rendering engines often repeat textures or geometries, leading to a model that overfits on a specific "style" rather than learning generalized features.
• The metadata integrity gap: Synthetic pipelines often lack the nuanced metadata (occlusion levels, pixel-perfect masks, and depth maps) required for complex tasks like 3D reconstruction or instance segmentation in cluttered environments.
  

While the research presented at WACV 2026 provides a visionary roadmap for the future of computer vision, the path to production is often less linear. For many engineers, the challenge isn't just generating synthetic data: it’s ensuring that data actually translates to real-world performance.

In conversation with a computer vision data expert: James Clough, VP of Engineering at Encord

To get a better sense of how teams attending WACV 2026 are navigating these hurdles on the ground, I sat down with James Clough, Encord’s VP of Engineering, to discuss the technical challenges of synthetic pipelines and how computer vision practitioners are bridging the gap between simulation and the field.

1. We’re seeing a massive spike in synthetic data papers at WACV 2026. From your perspective working with customers, what is the #1 reason teams are shifting to synthetic data right now? Is it just about saving money on labeling, or is something else driving this?

I don't think it's about saving money on labeling. I think it's more about growing faster and getting a larger scale of data than about saving money.

So, why is this happening now? Well, in the last few years, LLMs have grown very quickly and they've been able to scale by utilizing this really big free data set, which was all of the text on the internet to do pre-training. Then, the role of human generated data and human feedback was the post-training process where, after using lots of unlabelled data for pre-training, we then use smaller amounts of human labelled data for post- training to improve the performance of our language models.
But for other applications that aren't text, this same pattern has been harder because, if you think about an application like robotics for example, there isn't this huge free data set available to train your robot.

You don't have the equivalent of billions and billions of words and the understanding of which word should come next like you did for text available for these applications. So, the problem that you then have is, if you want to generate very large data sets for these other applications, you might have to do it all manually. And that is very expensive and very difficult - not just from a cost point of view, but from a time point of view. If you want to generate a billion hours of footage of video, then that's going to take you a long time.

So, the promise of synthetic data is that it allows you to generate really large data sets, much larger than you'd be able to do otherwise, and do so relatively quickly. You can still use human feedback in human generated data but as an equivalent to the post-training step that we saw in previous scaling

So, synthetic data is really there to generate scale which is needed to improve performance more than just as a cost cutting exercise and the reason it's happening now is because of the shift to non-textbased models also being trained at a very large scale.

2. When a team first switches to synthetic data, they often expect a 'perfect' model. What is the biggest 'hidden' challenge or risk they usually overlook during that first month of training?

I think that the biggest challenge with synthetic data is simply that it doesn't exactly capture the real world. Synthetic data is not the same as the real data. And that means there's a risk that it either misses some important scenarios or, in some important scenarios, it doesn't give your model the signal that you want. Your synthetic data is really only as good as the process you use to generate it.

In some applications, it can be relatively easy to generate synthetic data that matches the world you care about in an important way, but in other situations it's a lot harder.

A good example that people might not even think of as synthetic data is selfplay in games. So, when people were training AI to play chess, they didn't do it by playing lots of human chess players. They started out by training on old chess games that exist, but then you could also generate new data by playing chess synthetically against a computer. This meant you could train your system to get better and better using those games. That's generating synthetic data, in a way, and the reason it worked really well was that in chess the world is very easy to model. In fact, you can model it perfectly, right? From a computer's point of view, you can know absolutely everything that's going on on the chessboard at any given time. You have perfect information. It’s as good as real data.

But in other applications, like if you're trying to train a robot to clean your house, you could generate a synthetic virtual model of your house, like a computer game or a VR game, but it's not going to be exactly the same as your house. Everything isn’t going to interact in exactly the same way. You're not going to cover every single possibility that could happen and you're limited by how good your physics model of the world that you're shaping is. That's the risk.

3. WACV 2026 has a whole workshop (SAFE) dedicated to 'forensics' which is basically checking if data is 'good' or 'fake.' How can an engineer tell if their synthetic data is actually helping the model versus just teaching it to 'cheat' on unrealistic patterns?

I think there's probably two broad categories of things one can do here.

The first is to use classic interpretability tools. For lots of AI systems and machine learning models, there's a whole field of research in AI interpretability which is understanding what your system is doing and, in particular, the interpretability with respect to data. Understanding why your AI system performs in a certain way based on the data it was trained on.

You can apply those same techniques to synthetic data and understand whether a certain kind of synthetic data generated in a certain way causes your system to do X rather than Y.

The second approach is to consider that the reason we use synthetic data is that it's hard to get very large amounts of real data, and we need very large amounts of real data to do the training. However, you don't need very large amounts of real data to do the testing and do the evaluation.

You can still generate some real data at a much smaller scale and use it to do your evaluations. You might generate a hundred, or a thousand examples of real human annotated data and that might not be enough to train on, but it's enough to test whether your system does what you want it to at a relatively high degree of accuracy.

I think good practice is to use your real data for evaluation, and then you can assess with a high degree of accuracy whether your synthetic data is helping or not. You probably don't want to use your synthetic data for evaluation as well because, if there's a blind spot in your synthetic data, your system won't know how to do something and it also won't have appeared in your evaluations as well and you will have missed it in both cases.

4. If an engineer is looking at a massive pile of synthetic images and a smaller pile of real-world images, how does Encord Index or Active help them 'bridge' those two worlds so the model actually works in the field?

When synthetic images are used in training, it's very important to be able to retain the metadata associated with those images, which is to say, how did I synthetically generate them?
You'll have lots of different ways of generating synthetic data. You might have lots of different models or simulations you run, or different parameters of those simulations, and you want to know which ones produce good models or bad models. 

It's important to then have software that allows you to to index and curate all of that synthetic data so that you know which is which and where it's all come from. Encord Index is designed for this, and that way you can use the right data at the right time.

How could Encord help you do this? Well, one way of doing that would be the embeddings you use on Encord, which lets you visualize all of the data that you've used to train your system in one low-dimensional representation. You could visualize your real data and your synthetic data together. And then you might see, for example, that there's some region of your real data that is not covered by the synthetic data and that they overlap in some areas but not in others. This could tell you, perhaps, that there's some subset of my real data which is very different to my synthetic data. 

For example, all of your synthetic images of your robot in the house are taken during the daytime but some of your real images are taken at night. That would mean that, because you're not generating any synthetic images when it's dark, your robot might not be trained very well when it's dark. 

You'll be able to see that distinction in Encord Active by using that embedding suit.

5. If you were walking the floor at WACV 2026, what’s the one piece of advice you’d give to a computer vision team that is currently 100% dependent on synthetic data?

I think this relates to something I was saying before that synthetic data works very well when you have a good way of generating the world that your AI system will be interacting with. So, if the world your AI system will interact with is either quite simple, like the chessboard, or is just very easily understood with rules, then you have a good way of generating it and you can generate high quality synthetic data.
But, where you generate where you'll have problems is when the synthetic world and the real world diverge in important ways. It's probably always going to be the case that the synthetic data you generate won't be exactly the same as the real world. You're never going to make it absolutely perfect in most realistic scenarios, but that doesn't mean your synthetic data isn't useful.

Your goal should be to understand that divergence, understand in what ways your synthetic data is different from your real real world data. And, if you invest in understanding the answer to that question, then you will be able to build confidence that your synthetic data is similar to the real data in the ways that are important for your application. And, maybe it's different in some ways that are not important for your application.

I think the key thing to do would be to understand the ways in which your real data and your synthetic data are the same, and the ways in which they're different, and then ask yourself whether they're the same in the ways that are important for the AI system you're building.

The challenges James described aren't just theoretical risks, they are the primary hurdles that even the most advanced vision teams attending WACV 2026 face when scaling.

A standout example of overcoming this is Neurolabs, a leader in synthetic-first computer vision for the retail industry and an Encord customer. By looking at how they navigated the gap between their 3D asset generation and real-world supermarket shelves, we can see a blueprint for how companies attending WACV 2026 can use synthetic data successfully at scale. Read their case study here.