Rendering procedural textures for huge digital worlds

对虚拟3D场景中真实感和细节的要求不断提高，导致了大量的数据。一个主要的驱动力是纹理，它被广泛用于表示精细的视觉细节，如表面或位移的材料参数的变化。纹理通常是从非常适合创建随机纹理的过程模型中获得的，例如模仿自然现象。过程纹理是一种生成方法，其中纹理由一组函数来表示，这些函数被评估以产生最终的纹理。在这个项目中，我们构建了程序纹理图（PTG），它将生成过程表示为一个图，其中源节点是数学函数，内部节点是像素处理操作，汇聚节点是最终输出纹理。在典型的生产流水线中，纹理要么是预先计算的，这对存储要求非常高，要么是在纹理访问期间进行实时评估，导致许多冗余计算。我们的项目就是关于这个困境的。我们计划将程序纹理合成和照片级真实感渲染视为一个紧密耦合的实体，以使按需使用纹理合成渲染高度详细的场景成为可能-并通过新颖的缓存方案减少冗余计算，该方案考虑了从纹理评估到高质量渲染需求的管道的各个方面。后者需要纹理滤波，即计算由射线差分确定的纹理子区域中的纹素的加权平均值。我们计划通过首先推导一种新的纹理过滤理论来解决这种方法带来的沿着的众多挑战，该理论用于从PTG（存储颜色以及法线/位移）创建的纹理的预过滤和抗锯齿。关键是新开发的缓存算法，它利用了PTG提供的额外知识，即纹理如何从单个基函数和操作演变而来。过程纹理的第二个挑战是，尽管功能强大且灵活，但创建具有所需外观的纹理：对于艺术家来说，构建适当的图是乏味的，需要深入了解底层操作。我们的新方法需要向PTG添加技术元数据，这将使构建任务更加困难。为了克服这一点，我们希望通过半自动的方法从输入样本中促进图形的生产，这些样本通常用于生产渲染。我们将开发新的算法来提取一组基本功能和组合运算符的样本，以获得具有相似外观的图像。一个全自动的工具既不现实也不相关，因为艺术家需要控制结果。为此，我们的目标是一个反馈回路的方法：艺术家修复约束，而算法解决子问题。

The ever increasing demands on realism and detail in virtual 3D scenes lead to a tremendous amount of data. One major driving force is textures which are widely used to represent fine visual details, such as variation of material parameters across surfaces or displacements. It is common that textures are obtained from procedural models which are well suited to create stochastic textures, e.g. to mimic natural phenomena. Procedural texturing is a generative approach where textures are compactly represented by a set of functions which are evaluated to produce the final texture. In this project we build on Procedural Texture Graphs (PTGs), which represent the generative process as a graph where source nodes are mathematical functions, inner nodes are pixel processing operations and sink nodes are the final output textures.In a typical production pipeline, textures are either computed upfront which becomes extremely storage demanding, or are evaluated on-the-fly during texture accesses resulting in many redundant calculations. Our project is concerned with this quandary. We plan to treat procedural texture synthesis and photo-realistic rendering as one tightly coupled entity to make the rendering of highly detailed scenes feasible using texture synthesis on demand -- and reduce the redundant calculations by novel caching schemes accounting for all aspects of the pipeline from texture evaluation to the needs of high-quality rendering. The latter requires texture filtering, i.e. computing the weighted average of texels in a subregion of the texture determined by ray differentials. We plan to address the multitude of challenges that comes along with this approach by first deriving a novel texture filtering theory for prefiltering and antialiasing of textures created from a PTG (storing color as well as normals/displacements). The key will be newly developed caching algorithms which exploit the additional knowledge that a PTG provides, namely how a texture evolves from individual basis functions and operations.The second challenge with procedural textures is, although powerful and flexible, the creation of textures with a desired look: the construction of an appropriate graph is tedious for artists and requires in-depth knowledge of the underlying operations. Our new approach requires to add technical metadata to the PTG, which would make the construction task even more difficult. To overcome this, we want to facilitate the production of the graph by a semi-automatic approach from input exemplars, which are often given for production rendering. We will develop new algorithms to extract a set of elementary functions and combination operators from the exemplars to obtain images with similar appearance. A fully automatic tool would be neither realistic neither relevant, because artists need to control the result. To this end, we aim at a feedback loop approach: artists fix constraints while the algorithms solve sub-problems.