Learning Image Fractals Using Chaotic Differentiable Point Splatting
A. Djeacoumar, F. Mujkanovic, H.-P. Seidel, T. Leimkühler Max-Planck-Institut für Informatik, Saarbrücken, Germany · 2025
Read Learning Image Fractals Using Chaotic Differentiable Point Splatting by A. Djeacoumar, F. Mujkanovic, H.-P. Seidel, T. Leimkühler Max-Planck-Institut für Informatik, Saarbrücken, Germany (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-17230.txt.
Introduction
Adarsh Djeacoumar, Felix Mujkanovic, Hans-Peter Seidel, and Thomas Leimkühler teach a computer to find the fractal recipe hidden inside a picture, in a 2025 paper. Fractals repeat their patterns at every scale, which is why they model coastlines, ferns, and clouds so well. The inverse problem asks for the small set of transformations, the fractal code, that would regrow a given image and let it zoom endlessly.
Their algorithm tunes such codes with a custom fractal renderer whose output can be nudged by gradients, a method called differentiable point splatting. Random restarts and gradient steps together escape the many local traps of the search. Against rival inversion methods, theirs recovered cleaner codes and richer zoom-ins from a single image. The authors pitch the tool for graphics and for reading natural structure.
Catalog record
| CREATOR | A. Djeacoumar, F. Mujkanovic, H.-P. Seidel, T. Leimkühler Max-Planck-Institut für Informatik, Saarbrücken, Germany |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-17230
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning Image Fractals Using Chaotic Differentiable Point Splatting A. Djeacoumar, F. Mujkanovic, H.-P. Seidel, T. Leimkühler Max-Planck-Institut für Informatik, Saarbrücken, Germany Abstract Fractal geometry, defined by self-similar patterns across scales, is crucial for understanding natural structures. This work addresses the fractal inverse problem, which involves extracting fractal codes from images to explain these patterns and synthesize them at arbitrary finer scales. We introduce a novel algorithm that optimizes Iterated Function System parameters using a custom fractal generator combined with differentiable point splatting. By integrating both stochastic and gradient-based optimization techniques, our approach effectively navigates the complex energy landscapes typical of fractal inversion, ensuring robust performance and the ability to escape local minima. We demonstrate the method’s effectiveness through comparisons with various fractal inversion techniques, highlighting its ability to recover high-quality fractal codes and perform extensive zoom-ins to reveal intricate patterns from just a single image. \ConferencePaper \authorCGFStandardLicense Volume: 44 Issue: 2 ccs Computing methodologies Point-based models ccs Computing methodologies Rendering ccs Computing methodologies Machine learning † † year: 2025 † † year: 2025 † † editors: A. Bousseau and A. Dai † † editors-short: D. Ceylan and T.-M. Li † † editors-star: Y. Gryaditskaya and P. Memari † † editors-tutorial: K. Hildebrandt and R. Mantiuk † † editors-education: R. Kuffner dos Anjos and K. Rodriguez Echavarria † † editors-poster: T. Günther and Z. Montazeri † † editors-dc: L. Liu and G. Singh † † subject: EUROGRAPHICS CGF Vol No † † teaser: We introduce a novel method to recover a fractal description from an image containing a self-similar shape. Our hybrid optimization achieves state-of-the-art fractal inversion, enabling the synthesis of intricate details at any desired scale – illustrated here with 64x zoom-ins. 1 Introduction The geometry of nature is often fractal [ Man82 ] : Structures repeat themselves across different scales, forming self-similar patterns. A classic example is a fern, where each branch resembles both the overall fern structure and its smaller sub-branches (Fig. 1 ). Similar observations can be made for coastlines, terrains, trees, river networks, blood vessels, etc. Such intricate patterns can be explained by the repeated application of simple rules, referred to as fractal codes . Knowledge of underlying fractal codes is essential not only for describing and understanding complex systems [ Man82 , HBG∗95 , Tur97 , SHM05 , Man97 ] but also as an effective means of data compression [ Jac92 , Fis94 ] . In this work, we develop a novel method for extracting fractal codes from an image containing a self-similar pattern using modern differentiable point splatting [ KKLD23 ] . The fractal inverse problem – finding fractal codes that generate a given geometry – is a long-standing formidable challenge [ Bar88 ] and typically framed as an optimization problem [ Vrs91 , CLV∗95 , TCCC23 ] . This presents two key challenges. Figure 1: A fern exhibiting self-similarities. First, both the large-scale recursive application of fractal codes and the rendering of high-quality images are computationally demanding, especially within an iterative optimization framework. We observe that large-scale fractal synthesis, i.e., recursively executing fractal codes hundreds of thousands of times per image, and well-designed rendering, including proper anti-aliasing, are both essential for achieving high-quality inversion results. While computer graphics research has explored the efficient forward rendering of fractals over the past few decades [ Car80 , HD91 , Mar10 , dSNLV21 ] , there has been limited focus on improving efficiency in the inverse setting. To address this gap, we introduce a highly efficient parallel fractal generator, paired with a custom differentiable renderer. Specifically, we propose a model that optimizes the parameters of an Iterated Function System (IFS) [ Hut81 ] to produce a point set that is differentiably rasterized [ ZPVBG01 , KKLD23 ] . Our model enables scaling up fractal synthesis to a quality level that facilitates state-of-the-art fractal inversion results. The second key challenge in fractal inversion is the low dimensionality of the parameters, which results in highly non-convex objective energy landscapes that are notoriously difficult to navigate. To address this challenge, stochastic optimization techniques, such as evolutionary algorithms [ NG94 , QIG17 ] , have been employed. While these methods are effective at escaping local minima, they often follow inefficient optimization trajectories. In a different line of work, gradient-based optimization has been explored to solve the fractal inverse problem [ MP98 , VS99 , TCCC23 , BN24 ] . Gradients provide a strong guiding signal that enhances efficiency, but they are inherently prone to converging to local minima. In this work, we make the observation that leveraging the synergies between stochastic and gradient-based techniques significantly outperforms all previous methods in solving the fractal inverse problem. Our optimization employs an interleaved scheme that alternates between gradient descent [ KB15 ] and simulated annealing [ KGJV83 ] , ensuring stable optimization trajectories while maintaining the ability to escape local minima. Our approach recovers high-quality fractal codes from images. The key property of these codes is their ability to synthesize highly detailed patterns with virtually infinite resolution. These generative modeling capabilities allow for extreme zoom-ins on the original image, revealing intricate self-similarities and complex structures at any scale (Fig. ). We conduct extensive comparisons with both classical and modern fractal inversion techniques, demonstrating the clear superiority of our method. In summary, our contributions are: • A novel, highly efficient image fractal inversion technique based on differentiable point splatting. • An effective optimization routine that relies on synergies between stochastic and gradient-based optimization. • State-of-the-art fractal inversion results alongside an in-depth analysis of the proposed algorithm and its components. 2 Related Work 2.1 Fractals Fractals are complex geometric patterns that repeat infinitely at different scales, revealing self-similarity through recursive or iterative processes [ Man80 , Man82 , PJSF04 ] . Fractal geometry is not only crucial in many scientific disciplines [ HBG∗95 , Tur97 , SHM05 , Man97 ] , but it has also been applied to the study of natural images [ Pen84 , TPRC00 ] and visual art [ RHD08 ] , and is an indispensable tool for synthetic scene creation [ DHN85 , BJM∗88 , Ebe02 , TGSC24 ] . A wide variety of algorithms for creating fractals and self-similarities exists. Grammar-based techniques use formal rules to iteratively generate complex structures [ Lin68 , RS80 ] with diverse applications [ SRDT01 , WWSR03 , PL12 ] . Noise-based approaches combine random functions at different scales [ MVN68 , Kes82 ] . Escape-time algorithms generate fractals by iterating a function at each point until it escapes a boundary [ MEG04 , Jul18 ] . Furthermore, self-similarity can be achieved by tiling [ Ban91 , Fat01 , OCNG21 , CZHN22 ] . We build our approach on Iterated Function Systems, which generate fractals by repeatedly applying a set of geometric transformations [ Hut81 , BD85 , Bar88 , Elt87 ] . This method, which we review in Sec. 3 , is well-suited for point-based rendering [ TCCC23 , BN24 ] and has recently been shown to combine effectively with other fractal generation paradigms [ SK24 ] . Fractals have also been studied in 3D [ HSK89 , Nor82 ] , where efficient rendering is crucial [ Car80 , HD91 , Mar10 , dSNLV21