Terrain Diffusion — Infinite Panorama

Generate a seamless, arbitrarily-wide panorama with Stable Diffusion v1.5 and the infinite-tensor library.

This demo wraps the paper's self-contained annotated_infinite_panorama.py script: noise is sampled from a deterministic tiled Gaussian field, denoised in overlapping windows across several timestep phases (blended with a linear kernel), then VAE-decoded and cropped to the requested width. No outpainting/stitching artifacts — the panorama is coherent because every overlapping tile is denoised from the same underlying infinite noise field.

Paper: Terrain Diffusion (arXiv:2512.08309) · project page

Note: this Space demonstrates only the flat infinite-panorama demo from the paper's repository, not the full hierarchical planetary-terrain pipeline.

Prompt

Crop width (px)

Width of the final panorama. Larger = more overlapping tiles = slower.

512 2048

Inference steps

10 100

Guidance scale

1 15

Seed

Panorama

Examples

Citation

@inproceedings{goslin2026infinitediffusion,
      author    = {Goslin, Alexander},
      title     = {InfiniteDiffusion: Bridging Learned Fidelity and Procedural Utility for Open-World Terrain Generation},
      booktitle = {Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers},
      year      = {2026},
      pages     = {10 pages},
      publisher = {ACM},
      address   = {New York, NY, USA},
      doi       = {10.1145/3799902.3811080},
      url       = {https://doi.org/10.1145/3799902.3811080},
      series    = {SIGGRAPH Conference Papers '26}
}