Reproduce “Adversarial Learning of Classifier-Free Guidance Schedules” 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

Ashwini Pokle, Alexandre Galashov, Arnaud Doucet, Mauricio Delbracio, Valentin De Bortoli

The method trains a lightweight controller to adapt classifier-free guidance to each prompt, time step, and current noisy image, instead of using one global scale. A discriminator supervises the controller by matching guided samples to true noised-data marginals, with a CLIP reward also used for text-to-image generation. On MS-COCO 512×512, GAN+MC slightly improves alignment and preference over MMD+SC, but has worse FID; evidence is limited to two custom backbones.

Key insights

  • Enforcing marginal consistency through an adversarial framework provides a stable and effective signal for learning adaptive CFG scales.
  • Incorporating information from the noisy sample ($x_s$) allows for fine-grained, context-aware guidance, significantly enhancing text-to-image alignment.
  • A lightweight MLP generator leveraging summary statistics from the noisy latent space is computationally efficient and performs better than more complex architectures for this task.

Summary by alphaXiv.

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