Młynarczuk 和 Habrat (2025) 发表了一篇有趣的论文，介绍了使用生成式人工智能 (GenAI) 生成岩石微观图像的方法。研究探讨了生成对抗网络 (GAN) 和扩散模型 (Stable Diffusion) 在地质和采矿科学中的应用，用于基于现有训练数据集生成新数据。研究比较了不同架构和模型在文本到图像、图像到图像以及文本/图像到图像等场景下的表现，包括本地模型训练、基于预定义模型的迁移学习以及使用商业工具的迁移学习。结果表明，模型选择对图像质量有显著影响，生成的图像从与真实世界数据差异明显到与原始样本几乎无法区分的高分辨率表示。研究认为，生成合成数据可作为支持地质和采矿研究的工具，但需要仔细选择模型并正确配置其架构和参数。

🔬 研究主要关注使用生成式人工智能（GenAI）生成岩石微观图像，探索了生成对抗网络（GAN）和扩散模型（Stable Diffusion）等技术在地质和采矿科学中的应用。

💡 研究使用了多种方法来生成图像，包括文本到图像、图像到图像以及文本/图像到图像的转换。同时，研究也对比了本地模型训练、基于预定义模型的迁移学习以及使用商业工具的迁移学习等不同方式。

📊 研究结果表明，不同架构和模型的选择对生成的图像质量有显著影响，生成的图像质量差异很大，从与真实世界数据差异明显到与原始样本几乎无法区分的高分辨率表示。

⚠️ 研究强调，生成合成数据可以作为支持地质和采矿研究的工具，但需要仔细选择模型并正确配置其架构和参数。同时，研究也指出，由于地质材料的空间变异性和各向异性，机器学习（ML）或深度学习（DL）方法的结果可能存在显著的不确定性。

Interesting paper generating microscopic images of rocks using generative artificial intelligence (GenAI) published this week by Młynarczuk and Habrat (2025).

“The generation of synthetic images can be an important element in supporting the augmentation and analysis of multimedia data. It has applications in many scientific fields. Also, in geological and mining sciences.

This study presents generative artificial intelligence approaches, particularly on Generative Adversarial Networks (GANs) and diffusion models (Stable Diffusion), as widely used techniques for generating new data based on existing training datasets.

The performance of these algorithms and the results obtained both with and without transfer learning, using local resources as well as commercial solutions offering high resolution of the generated images, are presented. Results are presented from text to image, image(s) to image, and text/image(s) to image scenarios. Local model training, transfer learning based on a predefined model, and transfer learning using commercial tools were used.

The results indicate that the choice of architecture and model significantly influences the quality of generated images, ranging from visuals that differ from real-world data to high-resolution representations that are nearly indistinguishable from original samples.

As a result of this work, the possibilities of generating synthetic data as a tool to support geological and mining research were presented, considering the technological and practical aspects of implementing these solutions.”

Conclusion
“generating microscopic geological images of rocks is feasible; however, it requires careful selection of models and proper configuration of their architecture and parameters”, “Due to geomaterials’spatial variability and anisotropic properties, results obtained using machine learning (ML) or deep learning (DL) methods may carry significant uncertainty.” “need to combine statistical methods (ML) with approaches based on physical data and expert knowledge, enabling the proper development of advanced technologies. In the authors’opinion, artificial intelligence can bring significant innovations to geological and geotechnical engineering, but its development should not replace human labor and expertise.“

https://link.springer.com/article/10.1007/s12145-025-01934-6#Fig13