VLMedia (2014) is a rapidly growing, mobile-focused team reaching millions of smartphone users with social network projects, which are available on AppStore, Google Play, and Web platforms. Their main office is located at ODTÜ Teknokent (Ankara). VLMedia has more than 200 million registered users in more than 50 countries and supports 19 languages.
Technical/Scientific Challenge
The primary technical challenge addressed here involved implementing intricate style transfer techniques from 2D to 2D and 2D to 3D using neural networks. AI played a pivotal role in overcoming this challenge by enabling the creation of a robust hybrid method for style transfer. This innovative approach combined the Neural Neighbor Style Transfer (NNST) method with traditional style transfer, integrating AI’s deep learning capabilities. Leveraging AI’s neural network architectures, specifically the VGG19 model, allowed for advanced feature extraction and content-style balancing, crucial for achieving faithful and detailed style transformations across different dimensions. The hybrid method, by fusing style and content loss computations, ensured the preservation of target styles while retaining original content intricacies, significantly enhancing the realism and comprehensiveness of the style transfer process.
Solution
The challenge was addressed by developing a hybrid style transfer method, merging NNST with conventional techniques. Leveraging VGG19 (Visual Geometry Group consisting of 19 convolutional layers – https://viso.ai/deep-learning/vgg-very-deep-convolutional-networks/) model for feature extraction and content-style balancing, this method optimized style and content loss computations. Iterative adjustments fine-tuned processing times and weight ratios, achieving a balance between preserving target styles and retaining original content details.
Figure 1: Example Outputs
Figure 2: Benchmark Tests for Different Optimizers on TRUBA (Barbun-Cuda Nvidia P100 GPU)
Figure 2:Original and Stylized 3D Object
Business Impact
The current landscape demands intricate customization in virtual content creation. Harnessing HPC accelerates style transfer, expediting content personalization. This not only enhances competitiveness but also drives efficiency, enabling faster turnarounds and broader creative exploration in the industry.
Benefits
- Enhanced Customization: Tailored Virtual Reality (VR) environments boost user engagement and satisfaction.
- Streamlined Design Iterations: Rapid style transfer expedites design alterations.
- Reduced Production Timelines: Quick customization minimises lead times.
- Amplified Immersive Experiences: Personalized VR encounters deepen user connections.
Keywords
- Keywords: Neural Neighbor Style Transfer, DataParallel, Virtual Reality, Neural Networks
- Technology: HPC, AI
- Industry sector: Manufacturing & Engineering, VR Environments
- In the realm of Manufacturing & Engineering, the fusion of HPC and AI sparked a revolutionary leap in design customization. This integration expedited intricate design alterations, slashing lead times significantly. Moreover, envisioning this innovation within VR environments amplifies the potential impact, as Style Transfer techniques, powered by HPC and AI, promise to redefine customization possibilities in the immersive world of Virtual Reality. This success narrative highlights how HPC and AI, when embedded in VR environments, reshape workflows, fostering unparalleled design personalization and efficiency
Contact:
Hüseyin Bahtiyar, huseyin.bahtiyar@vlmedia.com.tr
Eda Yüksel, eda.yuksel@vlmedia.com.tr