Reproduce “Iterative Tilting for Diffusion Fine-Tuning” 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
Jean Pachebat, Giovanni Conforti, Alain Durmus, Yazid Janati
Iterative tilting turns reward-guided diffusion fine-tuning into sequential score updates, using reward values rather than reward derivatives. It gradually reweights samples and trains each updated score network with gradient descent. In a controlled two-dimensional Gaussian-mixture test, it qualitatively recovers the known reward-tilted target for 20, 50, 100, and 200 tilt steps. The evidence is limited to this synthetic problem, not real-world generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.