Bringing AI-generated images to enterprise use cases

Gretel's new image synthetics enable you to generate high-quality images at scale. Get started today with our free public preview and let us know what you think!

Introducing Gretel image synthetics

“Pics and it didn’t happen.”

For all the images that currently exist on the internet and in private databases, the right ones for specific tasks can be difficult to find, or simply don't exist. Imagine you’re a data scientist at McAgro Ventures, working on a machine learning model for resilient crop detection. You have a great library of plant pictures, and you can find lots of pictures of snow, but without a separate set of pictures of your plants in that snow, you’re missing critical data you need to train your model. Today, we're introducing Gretel's new image synthetics capabilities, to help you generate the images you need, when you need them.

Gretel image synthetics enables enterprises to harness image-focused generative AI, and compose their domain data with the knowledge built into foundation models. We designed our image synthetics system specifically for generating domain-specific images at scale, with an emphasis on the high-quality needed for training downstream ML models.

With a wide range of applications, from medical X-rays of rare conditions to simulations of different climates for smart agriculture, the future is synthetic.

How to generate images with Gretel

To introduce these capabilities, we’ll showcase how an auto insurance company might use image synthetics to improve their ability to accurately process broken windshield claims. Today, determining the cost of replacing a windshield requires expert knowledge and in-field assessment. Automating this process is challenging due to the strong data collection requirements. The insurance group may not want to collect customer images due to privacy constraints, and breaking the windshields of their own cars is quite undesirable.

Image synthetics are a potential solution. By training the model on publicly available images of broken windshields and publicly available images of car models, this group can combine the two concepts to produce unlimited high quality images of broken windshields to use in training a downstream damage pricing model. 

First, we'll train Gretel image synthetics to generate realistic images of broken windshields. Since we leverage a pre-trained model, we can start by using just 17 images of broken car windshields. This process is teaching your custom Gretel image synthetics model about the new concept of broken windshields. 

Figure 1: Synthetic broken windshields

We can generate the above using the following prompt, which includes a series of style cues to suggest realistic images. 

`broken windshield car front view, high quality photography, Canon Eos 5D, 100mm`

An auto insurer might need to be able to estimate damage pricing for broken windshields for specific makes and models, or just make sure that their model is reliable for particular years or styles of vehicle. To increase their model’s reliability or accuracy, they might need to generate images with particular features. Let’s do this ourselves by training Gretel image synthetics to generate realistic images of Acura cars by using 15 images of an Acura TSX Sedan 2012. We can use our prompt to control certain aspects like color!

`pink acura, front view, high quality photography, Canon Eos 5D, 100mm`

Figure 2: Synthetic Acura sedans. The pink color on the right was generated by prompting.

There were no pink cars in the training dataset! We combined a new color with our Acura concept to generate high-utility images by prompting the model with text input. The model already “knows” thousands of concepts, which means you can often synthesize the data you need even when you don’t have fine-tuning data. This synthetic data stays grounded in your data input and gains flexibility from the many concepts already in the model. 

Now the fun starts! While none of our Acura training images had a broken windshield, and none of our broken windshield training images were of an Acura, our training images helped familiarize the model with each of these concepts individually. It’s now able to combine both concepts by generating images of Acura cars with broken windshields. The best part is that thanks to the magic of image synthetics, we could do this without breaking a single windshield!

Figure 3: Synthetic red Acura with a broken windshield
Figure 4: Two further synthetic variations on an Acura with a broken windshield

Get started with Gretel image synthetics today

Gretel image synthetics is now available as a free public preview. We'd love to hear your feedback, whether you use it to build a custom use case, follow along with this windshield use case, or even create your own superhero avatar. Once available in Gretel’s synthetic data platform, the new image synthetics model will bring to images the same high-quality, scalable, and private synthetic data generation you're used to for tabular, natural language, and time series datasets.