Reproduce “NUWA-Infinity: Autoregressive over Autoregressive Generation for Infinite Visual Synthesis” 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
Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang and 1 more
Microsoft Research Asia and Peking University developed NUWA-Infinity, an autoregressive over autoregressive model that enables infinite visual synthesis, generating arbitrarily-sized high-resolution images and long-duration videos. The model achieved the creation of a 38912x2048 image with global consistency and outperformed prior AR-based models in various high-resolution generation tasks.
Key insights
- Hierarchical autoregressive modeling provides an effective way to manage both fine-grained local details and broad global coherence in arbitrarily-sized visual content.
- Caching localized, multi-layer hidden states is crucial for scaling autoregressive generation to very large outputs by maintaining context without prohibitive computational costs.
- Dynamic, order-aware positional embeddings are essential for flexible, multi-directional visual synthesis, allowing the model to adapt to various generation strategies beyond a fixed linear order.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.