ImageNet classification with deep convolutional neural networks
Alex Krizhevsky, Ilya Sutskever et al.
50
Citations
3
Influential Citations
—
Venue
2025
Year
Modularization is a cornerstone of computer science, abstracting complex functions into atomic building blocks. In this paper, we introduce a new level of modularization by abstracting generative models into atomic generative modules. Analogous to fractals in mathematics, our method constructs a new type of generative model by recursively invoking atomic generative modules, resulting in self-similar fractal architectures that we call fractal generative models. As a running example, we instantiate our fractal framework using autoregressive models as the atomic generative modules and examine it on the challenging task of pixel-by-pixel image generation, demonstrating strong performance in both likelihood estimation and generation quality. We hope this work could open a new paradigm in generative modeling and provide a fertile ground for future research. Code is available at https://github.com/LTH14/fractalgen.
This paper introduces a novel perspective on generative model design by borrowing the concept of fractals from mathematics. Instead of designing monolithic architectures, the authors propose to build generative models from atomic modules that are recursively composed, leading to self-similar structures. This modularization is a fundamental shift from current practices, which often treat generative models as fixed, hand-crafted architectures. The idea is analogous to the use of modularization in software engineering, where complex systems are built from reusable components, potentially making generative models more interpretable, scalable, and adaptable.
The paper's focus on pixel-by-pixel image generation is a challenging testbed that requires capturing long-range dependencies and fine-grained details. By demonstrating that fractal generative models can achieve strong performance in this setting, the authors provide evidence that their framework is not just a theoretical curiosity but a practical approach. This could influence how future generative models are designed, especially in domains where data is high-dimensional and structured.
The abstract reports strong performance in both likelihood estimation and generation quality, but does not provide specific numerical metrics. This is a limitation of the abstract; however, the fact that the paper has already received 50 citations suggests that the results are compelling enough to attract attention. The use of pixel-by-pixel generation is particularly demanding, and achieving strong results there indicates the method's robustness. Future work should compare against state-of-the-art autoregressive models and GANs to quantify improvements.
This work has the potential to reshape generative modeling by introducing a modular, recursive design principle. It could lead to more parameter-efficient models, as atomic modules can be shared across scales, and may enable easier scaling to larger datasets. The fractal approach also opens up new research directions, such as learning the recursive structure or adapting it to other modalities like text or audio. By providing a new paradigm, this paper invites the community to rethink how generative models are built and could inspire a wave of research on self-similar architectures.
Alex Krizhevsky, Ilya Sutskever et al.
Ashish Vaswani, Noam Shazeer et al.
Douglas M. Bates, Martin Mächler et al.
Diederik P. Kingma, Jimmy Ba