This article was sent and accepted as one of the top 5 best full-paper for the computing track* from SBGAMES 2021.

The Brazilian Symposium on Computer Games and Digital Entertainment is the most important Brazilian symposium in game development and research, covering several areas from computer graphics to artificial intelligence, storytelling research, among other topics. It is the second-biggest computer science conference in Brazil, only behind SBC's main event. (Note: SBC is the Brazilian Computing Society, a sort of equivalent organization to ACM, but in Brazil).

All articles must be written and presented in English.

This article continues a previous work from the university on cloud rendering that combined implicit and procedural modeling. By setting a limited amount of parameters, you could generate realistic-looking clouds.

However, it was kinda tough to generate good-looking clouds with it because you need to learn how the combination of its many parameters works until you find satisfactory parameter combinations for it.

So, our motivation was to use these clouds as a test case for a framework that could detect these parameters using an evolutionary process to generate realistic-looking clouds. Technically, instead of generating clouds, you could use a similar approach to other kinds of objects as well.

The problem is that we don't know how to describe what could be a realistic-looking cloud, so, we have used a convolutional neuron network to do that for us.


Geometric modeling has recently leveraged the power of Generative Design. Generative Design can be seen as a framework in which models can be generated by systematically exploring a space of shapes generated by recombination of parametric shape descriptors. In this work, we explore the use of Generative Design for modeling 3D clouds. Clouds are usually modeled in Computer Graphics as voxelized models by using a combination of implicit and procedural techniques. Hence, they are described by a large number of parameters. Although these parameters usually include parameters that define a combination of scalar fields, noise functions and affine transforms, the controlled use of such parameters is rather complex. How to tune them up to obtain a plausible result is not obvious. We propose a method based on generative design
combined with Machine Learning to produce families of cloud shapes automatically. Our generative design method is based on an evolutionary approach that generates instances of plausible 3D cloud shapes by optimizing a fitness function that measures the likelihood of a shape be a cloud. As the manual design of a fitness function is also quite complex, we propose using a Convolutional Neural Network to learn the fitness of arbitrary 2D views of the generated clouds. We perform several experiments that confirm the viability of the proposed method compared to manually modeled clouds.


-> Generative Design applied to Cloud Modeling (Article)

-> The model used for the fitness function to detect realistic clouds (for Pytorch).

* Slide 18/74, the top 5 nominations for best full papers for the computing track. This paper did not make the top 3.