In this work, we propose a new 3D-aware generative model with the design of an equivariant radiance field as the underlying 3D representation. With the translation-equivariant representation, our approach can produce infinite-scale 3D scenes by generating local scenes and then stitch them with smooth transitions and consistent concepts.

Thanks to the capacity of infinite synthesis, we can generate first-view driving video without any 3D prior or post-processing.

In this study, we introduce an innovative framework for scene generation, designed to create high-quality 3D scenes from textual descriptions. We recognize challenges in representation and supervision aspects and design essential methodologies to address them.

Hover the mouse over the image area to see the text prompt!

a spacious Dojo with a shogun. a washing room with a girl, in Japanese anime style.