README.md

metadata

title: Terrain Diffusion Infinite Panorama
emoji: 🏔️
colorFrom: blue
colorTo: red
sdk: gradio
app_file: app.py
pinned: false
short_description: Infinite-width panorama generation with SD1.5
preload_from_hub:
  - stable-diffusion-v1-5/stable-diffusion-v1-5

Terrain Diffusion — Infinite Panorama Demo

This Space demonstrates the infinite-width panorama generation technique from the Terrain Diffusion paper (arXiv:2512.08309, project page).

It wraps the paper repository's self-contained annotated_infinite_panorama.py script: Stable Diffusion v1.5 denoising is tiled across an unbounded 1D axis using the infinite-tensor library, so any crop of the panorama is generated from a single globally-consistent noise field rather than independently generated and stitched. This avoids the seams/repetition typical of naive tile-and-blend outpainting.

Scope note: only the flat infinite-panorama demo is wrapped here. The paper repository's hierarchical planetary-terrain diffusion pipeline and its Flask serving API are out of scope for this Space.

How it works

1. Deterministic tiled noise. A 1D-tiled Gaussian noise field is sampled such that any overlapping window of the (conceptually infinite) latent sees identical noise — the noise value at column x depends only on (seed, x // tile).
2. Phased denoising. The DDIM timestep schedule is split into phases at thresholds (400, 600, 750, 900). Each phase denoises overlapping latent tiles (stride 32, tile size 64) independently, then infinite-tensor blends overlapping outputs with a linear weight kernel before the next phase reads them back.
3. VAE decode. The fully-denoised latent is VAE-decoded tile-by-tile (pixel tile 512, stride 384) with the same overlap-blend trick.
4. Crop. The (infinite) decoded pixel tensor is sliced to the requested crop width and returned as a single image.

Using this Space

• Prompt: text description of the scene (applies uniformly along the panorama).
• Crop width: final output width in pixels (512–2048; default 1536). Larger widths require more overlapping tile evaluations and take longer.
• Inference steps: DDIM steps (default 40, paper default was 50).
• Guidance scale: classifier-free guidance weight (default 7.5).
• Seed: controls the underlying infinite noise field.

Model

Uses stable-diffusion-v1-5/stable-diffusion-v1-5 (fp16) as the base checkpoint. Note: the original runwayml/stable-diffusion-v1-5 repository was removed from the Hugging Face Hub; this Space uses the community mirror with identical weights, per current HF guidance.

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}
}