作者：Dana Cohen-Bar Elad Richardson Gal Metzer Raja Giryes Daniel Cohen-Or
最近在文本引导的图像生成方面的突破导致了文本三维合成领域的显著进展。通过直接从文本中优化神经辐射场（NeRF），最近的方法能够产生显著的结果。然而，这些方法对每个对象的位置或外观的控制是有限的，因为它们代表了整个场景。在需要在场景中细化或操作对象的场景中，这可能是一个主要问题。为了弥补这一缺陷，我们提出了一种新的全局局部训练框架，用于使用对象代理合成3D场景。代理表示对象在生成的场景中的位置，并可选择定义其粗略几何体。我们方法的关键是将每个对象表示为一个独立的NeRF。我们在优化每个NeRF本身和作为完整场景的一部分之间交替。因此，可以学习每个对象的完整表示，同时也可以创建具有风格和照明匹配的和谐场景。我们展示了u
Recent breakthroughs in text-guided image generation have led to remarkableprogress in the field of 3D synthesis from text. By optimizing neural radiancefields (NeRF) directly from text, recent methods are able to produce remarkableresults. Yet, these methods are limited in their control of each object’splacement or appearance, as they represent the scene as a whole. This can be amajor issue in scenarios that require refining or manipulating objects in thescene. To remedy this deficit, we propose a novel GlobalLocal trainingframework for synthesizing a 3D scene using object proxies. A proxy representsthe object’s placement in the generated scene and optionally defines its coarsegeometry. The key to our approach is to represent each object as an independentNeRF. We alternate between optimizing each NeRF on its own and as part of thefull scene. Thus, a complete representation of each object can be learned,while also creating a harmonious scene with style and lighting match. We showthat using proxies allows a wide variety of editing options, such as adjustingthe placement of each independent object, removing objects from a scene, orrefining an object. Our results show that Set-the-Scene offers a powerfulsolution for scene synthesis and manipulation, filling a crucial gap incontrollable text-to-3D synthesis.
论文链接：http://arxiv.org/pdf/2303.13450v1
更多计算机论文：http://cspaper.cn/