Reproduce “Breaking the Likelihood-Quality Trade-off in Diffusion Models by Merging Pretrained Experts” 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
Yasin Esfandiari, Stefan Bauer, Sebastian U. Stich, Andrea Dittadi
A plug-and-play sampling method merges two pre-trained diffusion models, one specialized for perceptual quality and the other for data likelihood, to resolve the inherent trade-off between these two metrics. This approach achieved improved Fréchet Inception Distance (FID) and Negative Log-Likelihood (NLL) rates on CIFAR-10 and ImageNet32 compared to individual models.
Key insights
- Accurate denoising at higher noise levels primarily shapes the global structure and overall perceptual quality of generated images.
- Data likelihood is highly sensitive to subtle, low-level pixel statistics, which are refined during low-noise denoising steps.
- Combining models specialized in these distinct aspects at appropriate stages of the denoising process can leverage their complementary strengths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.