Reproduce “PAGER: Progressive Attribute-Guided Extendable Robust Image Generation” on OpenResearch

About this paper

Zohreh Azizi, C.-C. Jay Kuo

The University of Southern California introduces PAGER, a non-deep learning generative model leveraging Successive Subspace Learning and Gaussian Mixture Models to produce images. PAGER achieves image generation quality comparable to deep learning methods, reduces training time by up to 325x on CPU, and maintains robustness to limited datasets, with only a 6% FID change on CelebA when training data is reduced by 20 times.

Key insights

• SSL's ability to transform images into a decorrelated latent space with approximately Gaussian features enables efficient and mathematically transparent distribution modeling using GMMs.
• A multi-stage, progressive generation pipeline effectively builds complex images from low-resolution initial outputs, enhancing stability and control.
• Non-iterative, data-driven methods like SSL and GMMs provide significant advantages in training speed and robustness to data scarcity compared to traditional deep learning paradigms.

Summary by alphaXiv.