TerrainDiffuser is a generative AI framework that enables dynamic texture generation for 3D terrains by leveraging text prompts and elevation data. Traditional terrain generation workflows often require manual sculpting and texturing, which can be time-consuming. TerrainDiffuser streamlines this process by generating high-quality, semantically consistent terrain textures directly from digital elevation maps (DEMs) and textual descriptions. This makes it a powerful tool for game development, simulation environments, and rapid prototyping.
- 🚀 Text-Guided Terrain Texturing – Generates terrain textures conditioned on user-provided text prompts (e.g., "snowy peaks with a dense forest below").
- 🏔 Elevation-Aware Generation – Ensures that textures are consistent with the terrain's height map, maintaining realism and structural coherence.
- 🔄 Flow Matching-Based Approach – Utilizes flow matching and multi-scale feature conditioning to balance geometric constraints with artistic flexibility.
- 🖼 Super-Resolution Post-Processing – Enhances generated textures using SwinIR to upscale outputs for high-resolution 3D rendering.
The following diagram illustrates the overall pipeline of TerrainDiffuser, from input elevation maps and text prompts to the final rendered 3D terrain. 3D rendering of the result is done using Blender.
- 🎮 Game Development – Rapidly prototype realistic terrain textures from simple sketches and text inputs.
- 🌍 Simulation & GIS – Generate accurate terrain representations for geospatial analysis.
- 🎨 Procedural Content Creation – Automate terrain design workflows for large-scale virtual environments.
Coming soon... Currently preparing for open use.
Coming soon? maybe?
This work was conducted as part of Project Research A at Simo-Serra Laboratory, Waseda University.