Research On Data-driven Fluid Animation Edit And Detail Enhancement

Fluid simulation is popular in CG movies,virtual reality and special effects.It has been a long-term hotspot on efficient and controllable generation of high-resolution fluid animation with rich details.Current fluid simulation methods based on physical equations suffer from complicated and time-consuming calculation costs,while methods based on deep learning technologies lack interpretable representations,which makes it a hard issue for ensuring edit and generation of fluids to be physically correct.For tackling the above difficulties,this thesis makes efforts on analysing multi-scale latent representations of fluid extracted by machine learning models such as convolutional neural networks and random forests.This thesis proposes a series of data-driven methods on fluid generation and detail enhancement.The core contributions of this thesis are:· A multi-scale bases decomposition method for fluids.For tackling the problem of lacking of interpretable fluid representation,the method maps multi-scale fluid features extracted by neural networks to a basis space which follows orthogonality constraint,incompressibility constraint and local support constraint.In addition,the method trains the bases decomposition network under the guidance of specific data sets so that further maps features to a local Eulerian interpretable space,which correlates multi-scale bases with physical parameters.· An editable fluid generation method based on the above interpretable bases.These bases together with physical parameters are utilized as inputs for a GAN to generate novel fluid velocity fields,which are later evaluated by a parameter evaluation discriminator on whether the generated velocity fields are coupled with the effects of physical parameters.In addition,temporal coherence constraint of fluid is considered when training GAN.Finally,the method succeeds in generating fluid animation under the control of multi-grained physical parameters rapidly.· A splash detail enhancement method based on random forest based distance.As generating splash details is time-consuming,the method proposes a metric of fluid patches based on random forests.Then the method retrieves the most similar candidate for low-resolution input patches from high-resolution spray patch repository.In addition,the method synthesizes retrieval results from different time steps to produce robust results.Finally,the method succeeds in generating splash details with high-resolution visual fidelity for low-resolution fluid simulation.· A super-resolution method based on convolutional neural networks for water surface waves.For tackling time-consuming high-resolution waves generation,the method customizes frequency-aware super-resolution CNNs for four dimensions as spatial resolution,temporal evolution,wave direction and wave number respectively.Then the method introduces a joint-training scheme to integrate results of the above CNNs.The method arrives at 13× speedup for 32× up-sampling of input scenes along all dimensions.The effectiveness and robustness of this thesis is validated by a variety of experiments.In addition,results of above methods are rendered with the help of rendering engines so that show realistic fluid animations.Finally,methods proposed by this thesis enriches fluid simulation technologies.