Károly Zsolnai-Fehér - Academia.edu (original) (raw)

Papers by Károly Zsolnai-Fehér

We present a learning-based system for rapid mass-scale material synthesis that is useful for nov... more We present a learning-based system for rapid mass-scale material synthesis that is useful for novice and expert users alike. The user preferences are learned via Gaussian Process Regression and can be easily sampled for new recommendations. Typically, each recommendation takes 40-60 seconds to render with global illumination, which makes this process impracticable for real-world workflows. Our neural network eliminates this bottleneck by providing high-quality image predictions in real time, after which it is possible to pick the desired materials from a gallery and assign them to a scene in an intuitive manner.
Workflow timings against Disney's "principled" shader reveal that our system scales well with the number of sought materials, thus empowering even novice users to generate hundreds of high-quality material models without any expertise in material modeling. Similarly, expert users experience a significant decrease in the total modeling time when populating a scene with materials. Furthermore, our proposed solution also offers controllable recommendations and a novel latent space variant generation step to enable the real-time fine-tuning of materials without requiring any domain expertise.

Real-time results of our method for simulating translucent materials (skin on the left, ketchup o... more Real-time results of our method for simulating translucent materials (skin on the left, ketchup on the right). Our separable subsurface-scattering method enables the generation of these images using only two convolutions (versus 12 in the sum-of-Gaussians approach [dLE07, JSG09]) and seven samples per pixel, while featuring quality comparable with the current state of the art, at a fraction of its cost. It can be implemented as a post-processing step and takes only 0.489 ms per frame on an AMD Radeon HD 7970 at 1080p, which makes it highly suitable for challenging real-time scenarios. Abstract In this paper we propose two real-time models for simulating subsurface scattering for a large variety of translucent materials, which need under 0.5 milliseconds per frame to execute. This makes them a practical option for real-time production scenarios. Current state-of-the-art, real-time approaches simulate subsurface light transport by approximating the radially symmetric non-separable diffusion kernel with a sum of separable Gaussians, which requires multiple (up to twelve) 1D convolutions. In this work we relax the requirement of radial symmetry to approximate a 2D diffuse reflectance profile by a single separable kernel. We first show that low-rank approximations based on matrix factorization outperform previous approaches, but they still need several passes to get good results. To solve this, we present two different separable models: the first one yields a high-quality diffusion simulation, while the second one offers an attractive trade-off between physical accuracy and artistic control. Both allow rendering subsurface scattering using only two 1D convolutions, reducing both execution time and memory consumption, while delivering results comparable to techniques with higher cost. Using our importance-sampling and jittering strategies, only seven samples per pixel are required. Our methods can be implemented as simple post-processing steps without intrusive changes to existing rendering pipelines.

Sophisticated global illumination algorithms usually have several control parameters that need to... more Sophisticated global illumination algorithms usually have several control parameters that need to be set appropriately in order to obtain high performance and accuracy. Unfortunately, the optimal values of these parameters are scene dependent, thus their setting is a cumbersome process that requires significant care and is usually based on trial and error. To address this problem, this paper presents a method to automatically control the large step probability parameter of Primary Sample Space Metropolis Light Transport (PSSMLT). The method does not require extra computation time or pre-processing, and runs in parallel with the initial phase of the rendering method. During this phase, it gathers statistics from the process and computes the parameters for the remaining part of the sample generation. We show that the theoretically proposed values are close to the manually found optimum for several complex scenes.

ACM SIGGRAPH 2012 Computer Animation Festival on - SIGGRAPH '12, 2012

We present a learning-based system for rapid mass-scale material synthesis that is useful for nov... more We present a learning-based system for rapid mass-scale material synthesis that is useful for novice and expert users alike. The user preferences are learned via Gaussian Process Regression and can be easily sampled for new recommendations. Typically, each recommendation takes 40-60 seconds to render with global illumination, which makes this process impracticable for real-world workflows. Our neural network eliminates this bottleneck by providing high-quality image predictions in real time, after which it is possible to pick the desired materials from a gallery and assign them to a scene in an intuitive manner.
Workflow timings against Disney's "principled" shader reveal that our system scales well with the number of sought materials, thus empowering even novice users to generate hundreds of high-quality material models without any expertise in material modeling. Similarly, expert users experience a significant decrease in the total modeling time when populating a scene with materials. Furthermore, our proposed solution also offers controllable recommendations and a novel latent space variant generation step to enable the real-time fine-tuning of materials without requiring any domain expertise.

Real-time results of our method for simulating translucent materials (skin on the left, ketchup o... more Real-time results of our method for simulating translucent materials (skin on the left, ketchup on the right). Our separable subsurface-scattering method enables the generation of these images using only two convolutions (versus 12 in the sum-of-Gaussians approach [dLE07, JSG09]) and seven samples per pixel, while featuring quality comparable with the current state of the art, at a fraction of its cost. It can be implemented as a post-processing step and takes only 0.489 ms per frame on an AMD Radeon HD 7970 at 1080p, which makes it highly suitable for challenging real-time scenarios. Abstract In this paper we propose two real-time models for simulating subsurface scattering for a large variety of translucent materials, which need under 0.5 milliseconds per frame to execute. This makes them a practical option for real-time production scenarios. Current state-of-the-art, real-time approaches simulate subsurface light transport by approximating the radially symmetric non-separable diffusion kernel with a sum of separable Gaussians, which requires multiple (up to twelve) 1D convolutions. In this work we relax the requirement of radial symmetry to approximate a 2D diffuse reflectance profile by a single separable kernel. We first show that low-rank approximations based on matrix factorization outperform previous approaches, but they still need several passes to get good results. To solve this, we present two different separable models: the first one yields a high-quality diffusion simulation, while the second one offers an attractive trade-off between physical accuracy and artistic control. Both allow rendering subsurface scattering using only two 1D convolutions, reducing both execution time and memory consumption, while delivering results comparable to techniques with higher cost. Using our importance-sampling and jittering strategies, only seven samples per pixel are required. Our methods can be implemented as simple post-processing steps without intrusive changes to existing rendering pipelines.

Sophisticated global illumination algorithms usually have several control parameters that need to... more Sophisticated global illumination algorithms usually have several control parameters that need to be set appropriately in order to obtain high performance and accuracy. Unfortunately, the optimal values of these parameters are scene dependent, thus their setting is a cumbersome process that requires significant care and is usually based on trial and error. To address this problem, this paper presents a method to automatically control the large step probability parameter of Primary Sample Space Metropolis Light Transport (PSSMLT). The method does not require extra computation time or pre-processing, and runs in parallel with the initial phase of the rendering method. During this phase, it gathers statistics from the process and computes the parameters for the remaining part of the sample generation. We show that the theoretically proposed values are close to the manually found optimum for several complex scenes.

ACM SIGGRAPH 2012 Computer Animation Festival on - SIGGRAPH '12, 2012