
Abstract With the recent surge in generative AI video models and neural rendering techniques, a key quality factor is the (spatio‐)temporal consistency of the generated or rendered sequence over its duration. Classic rendering methods, given the scene model is often readily available in full beforehand, inherently exhibit temporal and semantic consistency. This is frequently assisted for example by the appropriate antialiasing methods. Neural rendering and generative AI models, on the other hand, given rich training data, benefit from the generalisation capabilities of the advanced deep neural models to compensate for the lack of such scene attributes. The absence of a common scene model however can result in spatial and/or temporal inconsistencies in rendered sequences. This contribution aims to systematically categorise, review and compare the attributes of the surveyed methods specifically for modelling temporal coherence in neural rendering and generative AI. We group the relevant literature into six categories according to the principal mechanism through which temporal information is introduced, represented, reused or enforced. These categories cover temporal interaction within the model, temporally coherent representations, temporally coherent guidance, temporal context propagation and reuse, temporally structured diffusion noise, and lastly temporal losses, correction and smoothing. Each category is studied in detail regarding the main characteristics of the presented methods. We also provide a brief overview of the quality metrics employed by the reviewed methods for measuring temporal coherence for the related application areas. We conclude with a discussion on the future research avenues for modelling temporal coherence in this field.
Abstract In computer graphics and physics‐based animation, surface tension is an important factor in reproducing small‐scale fluid features. However, existing surface tension solvers often incur high computational costs and tend to become unstable under strong surface tension effects. To address this issue, we propose an SPH surface tension simulation method that combines geometric interface classification with an implicit solve. First, we construct a geometric criterion based on inter‐particle distances and neighborhood angles to identify liquid–air interface particles, liquid–solid interface particles, and bulk particles without relying on level sets or air‐phase particles. We further provide a threshold‐selection scheme independent of the particle radius. Building on this classification, we restrict the global pairwise cohesive interaction to an volumetric force applied only to interface particles and limit the degrees of freedom of the implicit system to interface particles. This design forms a pipeline that integrates particle classification, interface‐restricted cohesion solving, and pressure‐mediated force transfer. Experimental results show that our method significantly reduces computational cost, effectively reproduces surface‐tension‐induced fluid effects, and mitigates excessive local cohesion in canonical scenarios.
Abstract Power diagrams with prescribed capacities are widely used in geometry processing, including blue‐noise sampling, remeshing, and constrained clustering. The problem can be viewed as a semi‐discrete optimal transport problem and is commonly discretized by dense sampling into a capacity‐constrained assignment between n samples and m sites. Since this assignment is repeatedly solved in Lloyd‐type optimization, its efficiency is crucial to scalability. We propose an exact algorithm for the resulting discrete capacity‐constrained assignment problem. Unlike general‐purpose minimum‐cost flow solvers, which treat the input as an abstract graph, and entropic‐regularized approximations, which sacrifice exactness, our method exploits the asymmetry m ≪ n between sites and samples, performing augmentations on a compressed auxiliary graph with only m + 2 nodes while maintaining point‐level transfer costs using priority queues. We prove that augmenting paths in the full residual network can be projected to bounded‐length walks on the auxiliary graph, yielding an O(nm 2 log n) running time and near‐linear scaling in the number of samples for fixed m. Experiments on synthetic data and geometry‐processing tasks show that our method achieves one or two orders of magnitude speedup over classical algorithms while preserving exact optimality for the discrete transport problem, enabling efficient computation of capacity‐constrained power diagrams in large‐scale sampled settings .
Abstract 3D Gaussian Splatting (3DGS) remains challenging in dynamic outdoor scenes where dynamic distractors are coupled with large low‐texture regions such as roads and walls. Binary mask‐based foreground removal is insufficient—dynamic boundaries corrupt static optimization, while low‐texture regions lack reliable matching cues. In our method, we propose Reliability‐Aware Adaptive Gaussian Propagation (RAGP) for low‐texture dynamic outdoor scenes. Our method adapts patch scales based on local structure and static reliability for pixel‐level propagation, and fuses dynamic probability maps with rendering residuals into continuous reliability weights to constrain supervision, propagation filtering, and densification. Unlike foreground masking, RAGP converts dynamic priors into static reliability constraints and embeds them into geometric propagation. Experiments on dynamic outdoor benchmarks demonstrate that RAGP reduces artifacts and erroneous densification while improving low‐texture reconstruction and rendering fidelity.
Abstract Maintaining visual contact with moving targets in complex three‐dimensional environments is a fundamental challenge for aerial visual computing, as target visibility is jointly affected by scene geometry, dynamic motion, and viewpoint‐dependent occlusions. Existing aerial vision‐language navigation and tracking systems mainly focus on reaching static semantic goals or following non‐adversarial moving targets. However, in realistic pursuit scenarios, targets may actively evade the UAV, exploit buildings, trees, vehicles, and other environmental structures, and perform strategic maneuvers to break line‐of‐sight. To address this challenge, we introduce Adversarial Dynamic Target Visual Navigation, a new task paradigm that formulates UAV pursuit as a visibility‐aware viewpoint planning problem in dynamic 3D scenes. To support systematic evaluation, we construct FlySeek‐Bench, a graphics‐based benchmark featuring occlusion‐rich 3D environments, language‐conditioned target descriptions, rendered visual observations, and adversarial target policies that adapt to the UAV's motion. We further propose FlySeek, a unified visibility‐guided policy framework that combines predicted target hypotheses with semantic‐geometric visibility and occlusion reasoning for preemptive viewpoint selection. By coupling these action‐wise geometric priors with learned visual‐language preferences, FlySeek improves target observability against active evasion before visual contact is lost. Extensive experiments in simulated and real‐data‐reconstructed scenes demonstrate that FlySeek improves tracking robustness, visibility continuity, and pursuit success compared with representative AVLN, tracking, and planning baselines, especially under severe occlusion and adversarial target motion. These results highlight the importance of moving beyond passive target following toward visibility‐driven aerial pursuit in dynamic 3D environments. Code is available at https://github.com/JoshuaWenHIT/FlySeek .