Terrain Diffusion: A Diffusion-Based Successor to Perlin Noise in Infinite, Real-Time Terrain Generation

Abstract: For decades, procedural worlds have been built on procedural noise functions such as Perlin noise, which are fast and infinite, yet fundamentally limited in realism and large-scale coherence. We introduce Terrain Diffusion, an AI-era successor to Perlin noise that bridges the fidelity of diffusion models with the properties that made procedural noise indispensable: seamless infinite extent, seed-consistency, and constant-time random access. At its core is InfiniteDiffusion, a novel algorithm for infinite generation, enabling seamless, real-time synthesis of boundless landscapes. A hierarchical stack of diffusion models couples planetary context with local detail, while a compact Laplacian encoding stabilizes outputs across Earth-scale dynamic ranges. An open-source infinite-tensor framework supports constant-memory manipulation of unbounded tensors, and few-step consistency distillation enables efficient generation. Together, these components establish diffusion models as a practical foundation for procedural world generation, capable of synthesizing entire planets coherently, controllably, and without limits.

• The paper introduces Terrain Diffusion, a diffusion-based method that achieves infinite, real-time terrain generation with seed-consistency and hierarchical conditioning.
• It employs techniques such as InfiniteDiffusion, signed square-root transforms, and Laplacian encodings to enhance visual fidelity and reduce FID scores.
• The method supports parallel, order-invariant generation and integrates with engines like Minecraft, demonstrating practical real-world applications.

Terrain Diffusion: A Diffusion-Based Successor to Perlin Noise for Infinite, Real-Time Terrain Generation

Introduction and Motivation

Procedural terrain generation has traditionally hinged on noise functions such as Perlin noise, prized for their infinite extensibility, seed-consistency, and constant-time random access. These approaches, however, generate smooth patterns without realistic large-scale geospatial structure, failing to capture planetary coherence, hydrological features, or multi-scale diversity. With advances in generative diffusion modeling, the prospect of synthesizing terrain that exhibits natural structure, spatial coherence, and sharp details arises. However, canonical diffusion methods typically work on finite canvases and are computationally intensive, hampering their applicability for infinite, interactive world-generation tasks.

This work introduces Terrain Diffusion, which fundamentally repurposes score-based diffusion models to subsume the procedural noise paradigm. Through a novel InfiniteDiffusion algorithm, the method realizes unbounded, seed-consistent, and order-invariant synthesis at planetary scale, achieving real-time streaming with constant memory usage. Highlighted contributions include a hierarchical stack of diffusion and consistency models, novel data encodings for dynamic range stabilization, and an infinite-tensor framework enabling practical infinite-space sampling and manipulation.

Figure 1: A conceptual illustration of how InfiniteDiffusion leverages sliding windows to generate infinite terrain via deterministic, recursive region queries.

InfiniteDiffusion Framework

MultiDiffusion to Infinite Domains

Terrain Diffusion generalizes and extends MultiDiffusion methods to infinite domains by formulating the generation process over unbounded spatial coordinates. Each generation query involves a deterministic, finite set of local windowed region generations, assigning them explicit weights, and composing them into seamless output tiles. The recursion—tracing necessary dependencies across timesteps—halts when reaching the initial noise sample, which is deterministically seeded.

This approach admits several key formal properties:

• Seed-Consistency: Every queried region is a fixed, deterministic function of the initial seed and the coordinates, order-invariant and cacheable.
• Constant-Time Random Access: For any window-aligned region, the number of required model calls to generate from scratch is bounded independently of location, enabling discontinuous, on-demand exploration.
• Parallelism: All window updates at a fixed inference step are mutually independent and parallelizable within bounded computation rounds.

The practicality of the implementation is ensured with the Infinite Tensor framework—a library that efficiently manages infinite-domain tensor operations through tiling and dynamic streaming, providing the PyTorch API over virtually unlimited domains.

Data Representation and Model Hierarchy

Signed Square-Root Transform

Earth terrain displays extreme variance in elevation, confounding direct learning with unconstrained error magnitudes and unstable gradients. The paper applies a signed square-root transformation, $z \mapsto \mathrm{sign}(z)\sqrt{|z|}$, equalizing effective SNR across relief scales and focusing model capacity on small features including coastlines.

Figure 2: The signed-sqrt transform compresses standard deviation, distributing variance uniformly across mean elevations.

Hierarchical Laplacian Encodings

To further suppress low-frequency errors and reduce denoising requirements across long-range dependencies, Terrain Diffusion predicts Laplacian-encoded representations—downsampled smoothed low-frequency bands plus high-frequency residuals. Reconstruction decouples large-scale structure from fine detail, significantly reducing FID and increasing training stability and visual fidelity.

Multi-Stage Conditional Architecture

The procedural framework comprises a coarse planetary diffusion model for context generation, a core latent diffusion model that upscales to high-resolution tiles, and a final diffusion-based decoder that reconstructs accurate terrain detail from compressed latent spaces. Conditioning at each stage is imposed hierarchically, with explicit climate and geographic features available for precise user or procedural control.

Figure 3: Example of the initial input for hierarchical conditioning.

Figure 4: Twenty generated $1024 \times 1024$ terrain regions from a single world, covering a range of geological features with multi-scale coherence.

Experimental Results

Quantitative Metrics

The method is evaluated via FID-50k on non-tiled diffusion, non-tiled (distilled) consistency, and tiled (InfiniteDiffusion) generations, using central crops to mitigate window boundary artifacts. The Laplacian encodings and signed-sqrt pre-processing substantially reduce FID (core diffusion: 23.21 $\rightarrow$ 9.33, core consistency: 75.75 $\rightarrow$ 12.71), and InfiniteDiffusion introduces only minor additional FID degradation (final FID 17.87).

Latency analysis reveals that time-to-first-tile (TTFT) and time-to-second-tile (TTST) on a high-end GPU are highly competitive for practical streaming; even when traversing large worlds at high speed, tile generation remains under real-time constraints.

Qualitative and Integration Experiments

Qualitative grids of generated terrain demonstrate the model’s capacity for sharp, varied, and hydrologically consistent features without observable windowing or tiling artifacts. Integration with the Minecraft engine exemplifies real-world applicability: the model provides elevation and climatic biome values for seamless in-game world streaming using only the model’s outputs.

Figure 5: Nine Minecraft scenes generated from Terrain Diffusion, displaying high-fidelity landscape and climate-biome consistency.

Implications and Future Directions

The Terrain Diffusion system marks a transition from classical noise-based procedural synthesis toward generative AI foundations for infinite worldbuilding. The preservation of seed-consistency and random access ensures compatibility with existing gaming and simulation infrastructure, while allowing an unprecedented leap in visual and structural fidelity. By supporting hierarchical conditioning—including global climate and user control—the framework opens new possibilities for terrain authoring, interactive applications, and multi-modal guidance.

Algorithmic advances such as InfiniteDiffusion and the infinite-tensor API generalize beyond terrain, with immediate applicability to textures, environment maps, and any domain admitting local tiling and recursive sampling.

There are limitations. Global coherence ultimately requires external initialization of the coarse context model, typically achieved using classic noise or user-provided layouts. End-to-end learned models for the coarsest context remain challenging and are a promising direction for future research. Extending the hierarchy with additional environmental variables (soil, vegetation, etc.) and increasing resolution are anticipated directions that would further generalize the framework.

Conclusion

Terrain Diffusion establishes diffusion-based models as viable, high-fidelity, and practical successors to Perlin noise for real-time, unbounded terrain generation. Through the InfiniteDiffusion algorithm, infinite-tensor computation, and hierarchical conditioning, the system achieves seamless planetary synthesis with pre-specified randomness and order-invariant random access. These developments have significant implications for procedural content generation, simulation, and beyond, suggesting a broad applicability for score-based generative models in unbounded domains.