Reproduce “A Powerful Generative Model Using Random Weights for the Deep Image Representation” 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

Kun He, Yan Wang, John Hopcroft

The authors test whether reconstruction can work with an untrained VGG architecture by replacing trained weights with random values and optimizing pixels to match target activations. For texture synthesis, they balance layer contributions; ranVGG also selects weights against a reference. On two examples, ranVGG produces higher-quality reconstructions within 100 L-BFGS iterations than trained VGG after 2,000; this compares iterations, not runtime.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.