Left:
A. Synthetic MRI images of meningioma tumors generated using the Kidder model.
B. Synthetic MRI images of glioma tumors generated using the same model.

Right:
A. Actual training data from the CDD-CESM database (Kahled et al., 2022).
B. Synthetic mammography images generated by Kidder’s model based on text-to-image synthesis.

It has been shown that Kidder’s model generates synthetic medical images much faster than GANs, with GANs requiring 24–26 hours to train, while Kidder’s modified DreamBooth model only needed 10–15 minutes. Such text-input diffusion models can not only train diagnostic models that otherwise would not be trainable but also reduce the economic costs of collecting training data and the ethical concerns associated with using sensitive personal data for model training.

 

In the digital age, marketing is shifting from broad dissemination to personalized and precisely targeted strategies (Kim et al. 2022). Personalized marketing aims to recognize and address the individual needs of customers, including tailored emails, customized websites through cookies, recommendation engines, social media engagement, and finely tuned customer services (Dodds, 2024). By delivering personalized messages to their customers, companies can build deeper connections, increase engagement, conversion rates, and loyalty (Hovespian, 2024). At the core of massive personalization in modern marketing are AI systems that go beyond traditional data analysis, processing excessive amounts of user data to adapt strategies in real time to the behavior of each user (Akilkahkov, 2024).

Generative AI models have certainly earned their place in the industry by creating realistic images that appeal to customers, including in the media (Davenport and Mittal, 2022). But how effective are they? A study by TU Munich generated more than ten thousand images with seven modern generative text-to-image models and collected 254,400 human ratings (Hartmann et al., 2024). Respondents rated four of the seven tested models higher for “image quality,” with all seven models performing significantly better than human-created images. Considering the enormous cost difference between human freelancers and generative models, the clear advantage of the models becomes even more apparent: the researchers had to pay $100 for a single image from human freelancers, while with the same budget they could generate 2,500 images using DALL-E 3, the generative model with the highest scores.

Image generation process of the study by Hartmann et al.: A human-made image was converted into text, which was then fed into generative text-to-image models to create synthetic images.

Ethical concerns remain, however, as such generative AI models can be used to circumvent copyrights on training images (Avey, 2023). For example, if we follow the approach described above—converting a copyrighted image into a text representation, then using generative text-to-image models to transform the text into another image and using the synthetic image for marketing purposes—the original creator would practically never be able to detect the misuse of their creative work. Avey warns that most AI models take images from the internet without regard for whether they are copyrighted or not, as stated in the aforementioned report.

 

Pore structure refers to the general characteristics of the size, shape, distribution, and connectivity of the pore space (Jiang et al., 2007). Understanding its features and their impact on the physical properties of rocks, such as permeability, elasticity, and electrical properties, is important in many subfields of geosciences and petroleum engineering (Zhu et al., 2022). However, accurately characterizing the complexity of pore structure remains a challenge (Li et al., 2022), at least partly due to the heterogeneity of rocks, which makes visual examination in laboratory experiments difficult (Sun et al., 2017).

Digital rock modeling is an important method in geology for studying the microstructure and properties of rocks (Fang et al., 2020), which can be subdivided into the following subfields:

• Digital rock physics (DRP) can be used for the direct quantification of the structural and morphological parameters of rocks and for predicting flow properties at the pore scale (Sadeghnejad et al., 2021). DRP and its non-destructive methods have already become an important complementary method for reservoir characterization three decades ago (Blunt and King, 1991).
• Digital rock chemistry (DRC) is applied when changes in the pore structure are caused by interactions with dissolved substances (Sadeghnejad et al., 2021).
• Digital rock biology (DRB) is applied when changes in the pore structure are caused by microbial activities (Sadeghnejad et al., 2021).

Dense sandstone is a heterogeneous rock with diverse mineral compositions and multiscale pore structures, making it difficult to analyze using digital rock modeling (Chi et al., 2024). Chi et al. published in 2024 an attention-guided generative adversarial network that combines X-ray micro-computed tomography (Micro-CT) and scanning electron microscopy (SEM) images to create large-scale, high-precision rock images. Micro-CT images are non-destructive and cheaper but have lower resolution than the destructive and expensive SEM images. The goal was to develop a model capable of generating high-resolution synthetic SEM images from low-resolution Micro-CT images. Although they were not the first to use generative AI models to create high-quality rock images for digital rock physics (Niu et al., 2020; Chen et al., 2020), nor the first to combine Micro-CT and SEM images with generative AI models (Liu and Mukerji, 2022), they also addressed microstructures, including clay morphology.

Schematic representation of the attention-based GAN architecture by Chi et al.

The model is based on CycleGAN, where the model aims to improve low-resolution (LR) images to high-resolution (HR) while the adversary tries to reduce the resolution back to LR (Zhu et al., 2017). The authors used two mask generators, a content and an attention mask generator, in which higher-resolution images can be achieved by combining attention masks with their respective content masks and applying them to low-resolution images. The masks have the task of masking all parts of the image except for the small part that we want to draw the model's attention to, allowing the model to learn several characteristic features of different rock components simultaneously.

As intended by the authors, the model was able to generate synthetic high-resolution TEM images that preserve the micropores and trace minerals. A model that can create high-resolution synthetic images based on low-resolution real images is likely to provide effective technical support for digital rock modeling, characterization of pore structure, and numerical simulation of rock physics.