Our colleague Rob Reijtenbach recently investigated how generative AI can be used to create datasets for visual entailment. Visual entailment is the process by which AI models learn to understand whether an image supports, contradicts, or is neutral to a certain claim. This may sound complicated, but in practice it means that AI can better “understand” what is happening in images, and therefore respond smarter or make predictions.

The big problem in this field is that there are often too few good, annotated datasets to train AI models. Manually creating such datasets takes an incredible amount of time and money. Together with other experts, Rob demonstrates that generative AI can automatically create these datasets, allowing companies to develop models that understand visual information more quickly and cheaply.

This is groundbreaking for organizations because it makes AI projects more accessible. Companies can experiment with image-driven AI faster, develop new applications, and improve their products or services without huge investments in data collection. For example, organizations that want to ensure safety in the workplace can train AI to automatically check whether employees are wearing the correct protective clothing or whether safety signs are visible. This makes it possible to detect and mitigate risks more quickly, without needing hundreds of manual checks.

The research is published in the paper “Dataset Creation for Visual Entailment using Generative AI” by Suzan Verberne, Gijs Wijnholds, and thus Rob Reijtenbach. For organizations wanting to get started with AI, this offers a concrete way to get their projects off the ground faster.

Want to know more? Contact us to discover how generative AI can help your organization in training smart image models: info@mvrdw.nl. You can also directly send Rob a message via LinkedIn.

Link to article: https://arxiv.org/abs/2508.11605