High Frame Rate Photorealistic Flame Rendering via Generative Adversarial Networks

Abstract

In this paper we propose accelerating live rendering of flame using generative adversarial neural networks. The proposed method targets entertainment and simulation-based training industries whose demands for high fidelity and high frame rate increases steadily. The proposed approach takes image frames rendered with low voxel resolution (8 × 8 × 8 voxels at 90 FPS) and produces image frames equivalent to imagery produced from high voxel resolution (64 × 64 × 64 voxels) typically rendered at 3 FPS. The error was evluated using the structural similarity image metric (SSIM). The average error between generated image frames and the ground truth recorded 92:7%±4:6%.