InfiniteDiffusion: a diffusion model for generating infinite landscapes

A developer working under the name Xandergos has released InfiniteDiffusion — a system for generating landscapes for open-world games based on diffusion models.

The idea behind the project

The author sets out to combine two approaches that have traditionally been at odds in the game industry. On one side is procedural generation, which offers controllability and infinite scalability of the world but often falls short on visual quality. On the other are deep-learning-based methods, capable of producing more realistic and detailed textures and terrain but usually limited to a fixed output size.

InfiniteDiffusion proposes an architecture that lets a diffusion model generate landscapes without hard boundaries, while remaining practically usable in the procedural pipelines of game engines.

The release

The project is presented on a dedicated page where the author describes the technical implementation of the approach. The material sparked discussion in the Lobsters community, where users are curious about the details of how the model works and its applicability to real game projects.

The implementation details — including the diffusion-model architecture used and the approach to ensuring seamlessness across generated stretches of terrain — are laid out on the project page.

Source: Lobsters