Reproduce “Decentralized 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
David McAllister, Matthew Tancik, Jiaming Song, Angjoo Kanazawa
Decentralized Diffusion Models (DDM) proposes a framework for training large-scale diffusion models across independent compute resources by decomposing the learning problem. This approach demonstrated superior performance and efficiency, successfully enabling the training of a 24-billion parameter model using readily available hardware.
Key insights
- The marginal flow in diffusion models, representing the aggregate transport from noisy latents to data, can be decomposed into a linear combination of expert flows, enabling independent training of specialized models.
- Efficient inference can be achieved by using a router to select only the top-performing expert model for a given input, minimizing computational cost while maintaining high quality.
- Semantically coherent data partitioning is crucial for expert specialization, allowing individual models to learn more effectively on their assigned sub-distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.