Reproduce “No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion 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
Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, Romann M. Weber
Researchers from ETH Zürich and DisneyResearch|Studios introduce Independent Condition Guidance (ICG) and Time-Step Guidance (TSG) to address practical limitations of Classifier-Free Guidance in diffusion models. ICG enables CFG benefits for pre-trained conditional models without auxiliary training overhead, and TSG extends quality-boosting guidance to any diffusion model, including unconditional ones.
Key insights
- The unconditional score required by CFG can be effectively approximated from a purely conditional model's output by providing a statistically independent random condition.
- Internal time-step embeddings in diffusion models contain sufficient information to generate a powerful guidance signal that improves sample quality, even without explicit conditional inputs.
- Eliminating the auxiliary CFG training objective (condition dropping) allows conditional models to train more efficiently and achieve better final performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.