💻 computer science

HyperbolicDiffusion: Sharp & Scalable Tiled Generation on the Hyperbolic Plane

This paper introduces HyperbolicDiffusion, a training-free method that generates sharp, consistent, and reprojectable visual fields directly on the hyperbolic plane by employing a dynamic programming-based window placement strategy and a geometry-aware two-stage denoising process to resolve junction artifacts.

Original authors: Hugo Caselles-Dupré

Published 2026-08-05

📖 7 min read🧠 Deep dive

CC0 1.0

Original authors: Hugo Caselles-Dupré

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to paint a mural on a surface that doesn't behave like a flat wall or a round ball. In our everyday world, if you want to paint a huge picture, you usually stretch a canvas flat or wrap it around a sphere. But mathematicians have long been fascinated by a strange, infinite surface called the hyperbolic plane. Think of it like a piece of fabric that keeps getting wider and wider the further you go from the center, expanding so fast that the amount of space available grows exponentially. It's the kind of geometry that makes Escher's famous "Circle Limit" drawings possible, where fish get smaller and smaller as they swim toward the edge, yet the pattern never ends.

For a long time, computers could only generate images on flat screens or spheres. But what if we wanted to create a whole new world that lives directly on this weird, expanding surface? This is the challenge of diffusion models, a type of AI that creates images by slowly turning random static into clear pictures. Usually, these models work by chopping a big image into small, overlapping squares, fixing each square, and then blending them together. The problem is, when you try to do this on a hyperbolic plane, the "squares" get stretched and warped in crazy ways, and the edges where they meet often turn into a blurry mess. If you try to walk around this digital world, the blur gets worse, and the patterns break.

This paper introduces a new method called HyperbolicDiffusion that solves these problems without needing to retrain the AI. The researcher found a clever way to place their "painting windows" so they cover the infinite surface efficiently, and they built a two-step process to fix the blurry edges where the windows meet. The result is a sharp, consistent, and endlessly explorable hyperbolic world that looks just as good from any angle, like a digital version of Escher's impossible art that you can actually step inside.

The Problem: The Expanding Floor and the Blurry Edges

To understand why this is hard, imagine you are tiling a floor that is constantly expanding. In a normal room, you just lay down square tiles. But on a hyperbolic plane, the floor gets bigger so fast that if you want to cover a large area, you need exponentially more tiles. If you try to use a standard layout, you either waste millions of tiles or leave huge gaps.

Even worse, when you try to blend the images from these tiles together, the geometry fights back. Because the surface curves, a "square" window looks different depending on where it is and how it's oriented. When five different windows try to paint the same spot on the floor, they all see the neighborhood slightly differently. If you just mash their predictions together, the result is a ghostly blur where the edges don't line up. Previous attempts to do this on spheres (like the Earth) had similar issues, but on the hyperbolic plane, the problem is much harder because the area grows so fast.

The Solution: A Smart Tiling System and a Two-Step Fix

The author of this paper, working at Obvious Research in Paris, came up with a solution that doesn't require teaching a new AI from scratch. Instead, they built a smart system to manage how the existing AI paints the hyperbolic world. Their method, HyperbolicDiffusion, works in three main parts.

1. The Smart Tiling (Hyperbolic Blooming Cover)
First, they needed a way to place the "windows" (the squares the AI looks at) so they cover the entire area without wasting resources. They invented a method called the Hyperbolic Blooming Cover (HBC). Imagine planting flowers in a garden that keeps getting bigger. Instead of planting them randomly or in a rigid grid, HBC plants them in perfect rings that grow outward. It uses a clever math trick (dynamic programming) to figure out exactly how many flowers are needed in each ring to cover the ground perfectly.

The paper shows that this method is incredibly efficient. It can plan a layout with over 11 million windows in just 5.44 seconds. More importantly, they proved that for the areas they tested, their layout is never more than 10% away from the absolute best possible arrangement. This means they aren't wasting thousands of extra windows just to be safe; they are using almost the minimum number required.

2. The Shared Canvas (Stage A)
Once the windows are placed, the AI starts painting. The key here is that every point on the hyperbolic surface has a permanent ID, like a unique address. Even if a point appears in five different windows, it's always the same "address." The AI looks at the overlapping windows, makes a guess at what the image should look like for each, and then blends those guesses together based on how much it trusts each window. This creates a single, unified image on the hyperbolic surface.

However, the author found that this first step isn't perfect. Because the windows are warped differently, the AI sometimes paints slightly different versions of the same detail in the overlapping areas. This creates a "ghosting" effect where edges look blurry or doubled, especially in complex scenes like buildings or forests.

3. The Repair Crew (Stage B)
This is where the second part of their magic happens. They realized that they could predict exactly where the blur would happen just by looking at the geometry, without even needing to see the image first. They calculated a "multiplicity mask"—a map that shows where multiple windows are fighting over the same spot.

Where this map shows a conflict (where two or more windows are contributing equally), they run a second stage of repair. They take those specific blurry spots, add a little bit of noise back in (like shaking up a jar of sand), and ask the AI to repaint them again, this time focusing only on making the edges sharp and consistent. This "re-noising" step fixes the ghosting and makes the details crisp.

What They Found and What It Means

The researcher tested their method on everything from simple textures like clouds to complex scenes like indoor rooms and architectural frescoes. They found that:

The Tiling Works: Their HBC method successfully covered areas with radii up to 12 (a massive scale in hyperbolic terms) using the predicted number of windows, and the planning was always fast (under 6 seconds even for the largest layouts).
The Blur is Gone: In their tests, the second repair stage increased the sharpness of the edges significantly. They measured the "gradient magnitude" (a way to measure how sharp edges are) and found it improved by an average of 15%, and the "Laplacian magnitude" (another sharpness metric) improved by 17%.
It's Not Just a Flat Image: Unlike other methods that generate a flat picture and then warp it into a circle, this method generates the image directly on the hyperbolic surface. This means you can move a "camera" around the world, and the image stays sharp and consistent. A pattern that looks tiny at the edge will look huge in the center, but it won't get blurry or break apart.

The paper explicitly rules out the idea that a single-stage process (just painting once and blending) is enough for complex, structured content. They tried several one-stage variations, like using denser windows or blending ownership differently, but none of them could remove the blur without losing detail. The two-stage approach was the only one that worked.

They also showed that this method can be adapted to spheres (like the Earth) and outperforms previous methods there too, suggesting the idea of geometry-based repair is powerful.

Why This Matters

This work opens the door to creating "navigable hyperbolic murals" and non-Euclidean game worlds. Imagine a video game where you can walk through a hallway that gets infinitely wider the further you go, or a VR experience where you can explore a cathedral that expands forever. Because the image is generated directly on the curved surface, you can move your viewpoint, and the world remains sharp and coherent. It's a step toward generative art that truly understands the strange, beautiful geometry of the hyperbolic plane, turning a mathematical curiosity into a playground for digital creativity.

Technical Summary: HyperbolicDiffusion

1. Problem Statement

Existing diffusion models for large-format images, such as MultiDiffusion, operate on a flat Euclidean canvas by denoising overlapping rectangular crops. While effective for planar images, this approach cannot natively generate images that "live" on the hyperbolic plane ( $H^2$ ). Simply generating a flat image and warping it into a Poincaré disc post-hoc limits the utility of the representation; it prevents interactive navigation where motifs can expand from the rim to the center without regenerating content.

Direct generation on $H^2$ presents two unique challenges:

Exponential Scaling: Unlike the finite sphere ( $S^2$ ), the area of a hyperbolic disc grows exponentially with its radius ( $A(D_R) = O(e^R)$ ). Standard tiling strategies that waste a small fraction of views on a plane or sphere become prohibitively wasteful on $H^2$ , potentially requiring millions of windows.
Junction Artifacts: Previous attempts to share latent variables across curved manifolds (e.g., SphereDiff) suffer from "blur" or "ghosting" at multi-window junctions. This occurs because different views of the same point see differently warped neighborhoods due to curvature, causing the denoiser to produce inconsistent local features that blend poorly.

2. Methodology

The author proposes HyperbolicDiffusion, a training-free method that generates finite visual fields directly on $H^2$ using a two-stage process.

2.1. Hyperbolic Blooming Cover (HBC)

To address the exponential scaling problem, the paper introduces HBC, a training-free algorithm for optimal window placement.

Objective: Cover a target hyperbolic disc $D_R$ with the minimum number of "trusted" discs (radius $r_t$ ) such that every point in $D_R$ lies within at least one trusted region.
Algorithm: HBC exploits the radial symmetry of $H^2$ . It models the problem as a one-dimensional weighted interval cover on the radius $[0, R]$ . By enumerating feasible rings of window centers and solving this via dynamic programming, HBC generates a layout that "blooms" outward.
Efficiency: The algorithm outputs a compact list of ring parameters (population, radius, phase) rather than storing millions of individual coordinates. It runs in seconds (e.g., 5.44s for 11.4 million windows) and provides theoretical guarantees that the window count is within 10% of the optimal solution for $r_t \geq 1.151$ .

2.2. Stage A: Shared Latent Fusion

This stage generates the initial field using a standard frozen diffusion model (UNet) and a VAE.

Permanent Surface IDs: Every point on the hyperbolic surface is assigned a permanent latent ID. These IDs form a shared latent canvas.
Denoising: Windows gathered by the HBC layout map local surface points to the model's native square array. The model denoises these local views.
Fusion: Predictions for the same surface ID from overlapping windows are fused in latent space using a weighted average (similar to MultiDiffusion).
Limitation: While this ensures consistency of the latent variable, it does not guarantee that the neighborhoods seen by different windows are identical due to curvature. This leads to residual blur at high-multiplicity junctions where multiple windows contribute.

2.3. Stage B: Geometry-Attached Junction Repair

To eliminate the blur and ghosting caused by curvature-induced disagreement, a second stage is applied.

Mask Generation: A mask is generated purely from geometry, not image content. It calculates the "effective multiplicity" ( $m_{eff}$ ) of each point. Points where $m_{eff} \geq 2$ (indicating significant contribution from multiple competing windows) are flagged for repair.
Repair Process:
1. The masked regions are covered by larger surface windows.
2. The Stage A image is re-encoded into the latent space.
3. Scheduler noise is added at a strength of 0.55.
4. The model re-denoises only the masked regions using the original prompt.
5. A single "owner" window writes the repaired values back to the $H^2$ field.
Result: This process repairs the specific regions where geometric distortion caused the denoiser to disagree, restoring sharpness without re-generating the entire field.

3. Key Contributions

Hyperbolic Blooming Cover (HBC): A dynamic programming approach that solves the window placement problem on $H^2$ in seconds with near-optimal efficiency, handling the exponential area growth of the hyperbolic plane.
Two-Stage Generation Mechanism: The identification that shared latent IDs alone are insufficient for sharp hyperbolic generation. The proposed two-stage pipeline (Shared Fusion + Geometry-Attached Repair) is empirically shown to be the only mechanism capable of removing junction blur while maintaining large compositions.
Training-Free Approach: The method utilizes existing, frozen diffusion models and VAEs, requiring no retraining or fine-tuning of the generative backbone.
Native Hyperbolic Representation: Unlike prior work that warps flat images, this method generates fields natively on $H^2$ , enabling true navigability where content scales naturally with camera position.

4. Results

Scalability: HBC successfully generated layouts for radii up to $R=12$ (involving millions of windows) with zero holes and window counts within 10% of the theoretical optimum. Planning time remained under 6 seconds even for the largest configurations.
Image Quality: In qualitative evaluations involving structured content (architecture, frescoes), Stage B significantly reduced blur and ghosting. Quantitative metrics showed a 15% increase in masked mean gradient magnitude and a 17% increase in masked Laplacian magnitude compared to Stage A alone.
Comparison to SphereDiff: When applied to structured indoor scenes (a harder test than textures), the proposed method outperformed the SphereDiff baseline, which exhibited washed bands and blur at junctions.
Interactivity: The system supports interactive navigation (e.g., WebGL prototypes) where the camera moves across the field, reprojecting the same underlying $H^2$ data without regeneration.

5. Significance and Claims

The paper positions HyperbolicDiffusion as a generative counterpart to M.C. Escher's Circle Limit series, providing a method to create sharp, consistent, and navigable hyperbolic murals and environments.

The author modestly claims that their method solves the specific technical bottlenecks of exponential placement cost and junction quality failure on the hyperbolic plane. They emphasize that the two-stage mechanism is a necessary innovation to handle the geometric disagreement inherent in curved manifolds. The work does not claim to solve all manifold generation problems but specifically targets the generation of finite visual fields on $H^2$ with high fidelity and scalability. Future work is suggested to tighten finite certificates and explore single multiresolution diffusion states.