Reproduce “Fractal Generative Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianhong Li, Qinyi Sun, Lijie Fan, Kaiming He
Researchers from MIT CSAIL and Google DeepMind introduce Fractal Generative Models, a framework that recursively composes generative models into self-similar architectures. This method achieves state-of-the-art likelihood on pixel-by-pixel image generation for ImageNet, producing high-quality images with enhanced computational efficiency.
Key insights
- Fractal-like, self-similar architectures offer an effective and computationally efficient means to model complex, hierarchical data distributions directly at the pixel level.
- Modularizing entire generative models as atomic building blocks enables the construction of intricate generative systems, better suited for data lacking inherent one-dimensional sequence order.
- Direct pixel-by-pixel generation, when structured hierarchically with a fractal design, can achieve state-of-the-art likelihoods and high-quality outputs without the reconstruction errors introduced by latent space tokenizers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.