9:45 AM - 10:00 AM
[HGM04-04] DEMFLUX Geology-to-Terrain Model: Automatic Terrain Generation by Generative AI Trained on Actual Topography
Keywords:Large-scale topography, Plate Tectonics, Generative AI
Automatic terrain generation has a long history, dating back to early procedural methods such as Perlin noise and fractal algorithms in the late 20 centuries. While software like Terragen was popular at the time, these approaches often produced repetitive or fractal patterns and neglected key geological processes, particularly erosion and the diversity of rock types, faults, and volcanic features. Consequently, these early techniques lacked realism. Although the introduction of river drainage simulations in the 2010s provided slight improvements, the outputs still fell short of capturing geological consistency. In the 2020s, however, neural network–based approaches started to enable the generation of more diverse and geologically informed terrains, in contrast to the limitations of earlier procedural methods.
In this study, I developed large-scale (2,000 × 2,000 km and 4,000 × 4,000 km) terrain models using a geologically trained text-to-image diffusion model. This work builds upon FLUX (Black Forest Lab, 2023), which was fine-tuned with the GEBCO 2023 Grid (GEBCO, 2023) and my own geological annotations. The training dataset consisted of 644 grayscale images of actual Earth topography, each paired with a corresponding text file describing the geological context. The dataset was augmented through systematic rotations of the original topographic images, increasing sample diversity.
Qualitative evaluation suggests that the resulting terrain captures essential aspects of plate tectonics, including the bimodal distribution of oceanic and continental crust, oceanic ridges, transform faults, subduction zones, and orogenic belts. The generated terrains realistically transition among convergent, divergent, and strike-slip boundaries. However, resolution constraints and limited training data occasionally hinder the model’s ability to generate finer-scale features such as volcanic fronts, accretionary prisms, and complex interactions at triple junctions. Additionally, flat plains sometimes exhibit noise, likely due to the dynamic range constraints of 8-bit grayscale images.
Overall, this diffusion-based model demonstrates promise for realistic and geologically coherent terrain generation. It also highlights the potential of linking geological terminology directly to resultant topographic forms. Future work will focus on integrating additional constraints, such as those provided by ControlNet (Illyasviel, 2022), to allow greater control over features like plate boundary geometry, which could be particularly useful for paleogeographic reconstructions.
In this study, I developed large-scale (2,000 × 2,000 km and 4,000 × 4,000 km) terrain models using a geologically trained text-to-image diffusion model. This work builds upon FLUX (Black Forest Lab, 2023), which was fine-tuned with the GEBCO 2023 Grid (GEBCO, 2023) and my own geological annotations. The training dataset consisted of 644 grayscale images of actual Earth topography, each paired with a corresponding text file describing the geological context. The dataset was augmented through systematic rotations of the original topographic images, increasing sample diversity.
Qualitative evaluation suggests that the resulting terrain captures essential aspects of plate tectonics, including the bimodal distribution of oceanic and continental crust, oceanic ridges, transform faults, subduction zones, and orogenic belts. The generated terrains realistically transition among convergent, divergent, and strike-slip boundaries. However, resolution constraints and limited training data occasionally hinder the model’s ability to generate finer-scale features such as volcanic fronts, accretionary prisms, and complex interactions at triple junctions. Additionally, flat plains sometimes exhibit noise, likely due to the dynamic range constraints of 8-bit grayscale images.
Overall, this diffusion-based model demonstrates promise for realistic and geologically coherent terrain generation. It also highlights the potential of linking geological terminology directly to resultant topographic forms. Future work will focus on integrating additional constraints, such as those provided by ControlNet (Illyasviel, 2022), to allow greater control over features like plate boundary geometry, which could be particularly useful for paleogeographic reconstructions.