Research Radar

arXiv cs.GR details a new approach for masking judder in foveated rendering by phase-shifting synthesized noise instead of rerendering or blurring the image. The paper frames the problem around low temporal rates, where motion in the periphery can become visibly stuttery even when spatial detail …

• A paper proposes masking judder in foveated rendering with phase-shifted synthesized noise.
• Ville Cantory, Danhua Zhang, Victoria Interrante, and Piotr Didyk.

arXiv cs.GR details WorldBench, a benchmark and judge for open-ended Three.js voxel worlds generated by large language models. The core idea is simple: if a model can write an interactive 3D world, evaluation should inspect the running scene, not just a handful of renders or …

• WorldBench is a benchmark and judge for LLM-generated Three.js voxel worlds
• Krish Bakshi; evaluated Claude Fable 5.1, GPT-6 Astra, Kimi K3, Grok 4.7, and Gemini 3.1 Pro

arXiv cs.GR details GaussianBench, a physics-fidelity evaluation suite for physics-integrated Gaussian scene representations. The goal is to move beyond “looks right” rollouts and test whether a system’s internal mechanics actually hold up when scenes are pushed by new forces, material edits, contact, or thermal effects. …

• GaussianBench is a physics-fidelity benchmark for physics-integrated Gaussian scene representations.
• Author: Chukwudalu Dumebi-Kachikwu; systems tested include PhysGaussian, GaussianFluent, OmniPhysGS, PhysDreamer, Physics3D, and GASP.

arXiv cs.GR details FeatureZ, a new lossy compression framework for structured volumetric scalar fields that aims to preserve derived geometric and topological features, not just raw sample values. The work targets data used in scientific visualization, where losing an isosurface, merge tree, or Morse-Smale complex …

• FeatureZ is a feature-preserving lossy compression framework for structured volumetric scalar fields.
• Nathaniel Gorski, Xin Liang, Hanqi Guo, and Bei Wang submitted the paper to arXiv cs.GR.

arXiv cs.GR details RiCo, a neural simulation approach for rigid-body interactions that narrows cross-object reasoning to sparse neighborhoods of contact surface points. For game teams building physics-heavy systems, the interesting part is the shift from whole-scene modeling to local contact reasoning, which is closer to …

• RiCo is a neural rigid-body simulation method built around local contact reasoning.
• Ruixiang Ouyang and six coauthors; published on arXiv cs.GR.

arXiv cs.GR details OuroWorld, a new framework for turning static 3D Gaussian Splatting scenes into endlessly looping 3D cinemagraphs. The pitch is straightforward for real-time and content teams: take an explorable 3D scene and make it feel alive, with motion that stays coherent from different …

• OuroWorld turns static 3D Gaussian Splatting scenes into looping 3D cinemagraphs.
• Paper by You-Zhe Xie, Ting-Wei Chou, Yu-Hsuan Li, Kaipeng Zhang, Zhixiang Wang, and Yu-Lun Liu.

arXiv cs.GR details SpaceFlow, a training-free pipeline for locally controllable 3D generation that combines text prompts with a set of geometric primitives. Each primitive acts as a proxy for an object part, and each part can be assigned its own control level so some regions …

• SpaceFlow is a training-free pipeline for locally controllable 3D generation from text and geometric primitives.
• Authors include Neil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov, Felicia Scharitzer, Marc Pollefeys, Ata Celen, Sayan Deb Sarkar, and Elisabetta Fedele.

arXiv cs.GR details Tabula Rasa, a Monte Carlo method for generating time-varying Gaussian noise with controlled variance and temporal correlation. The paper frames the problem as joint Monte Carlo estimation of pixel reconstruction and variance, borrowing the idea of sketching from database research. The practical …

• Tabula Rasa proposes Monte Carlo estimation of unit-variance noise with controlled spatio-temporal correlation.
• Authors include Tobias Ritschel, Yang Zhou, Nick Milef, Mikhail Dereviannykh, Chen Liu, Christophe Hery, and Carl Marshall.

arXiv cs.GR details Neural Caching of Prefiltered Radiance for Specular Lighting, a neural radiance caching variant aimed squarely at specular transport in real-time path tracing. The method uses a reflection-direction parameterization plus a roughness-dependent radiance target, letting the network predict prefiltered incoming radiance at surface …

• A neural radiance caching variant for specular lighting and real-time path tracing
• Dmitrii Klepikov and Vladimir Frolov; arXiv cs.GR

arXiv cs.GR details SubDGuide, a modeler-inspired agentic workflow for reconstructing subdivision surfaces from dense meshes. The key idea is to let a planner reason about cage resolution, feature mapping, and broad-form fitting, while geometry tools handle every actual vertex or connectivity change. That separation matters …

• SubDGuide is an agentic workflow for mesh-to-SubD reconstruction.
• Authors: Mengnan Jiang, Christian Franke, Michele Franco Adesso, Antonio Haas, Grace Li Zhang.

arXiv cs.GR details Fluid-Gen-Zero, a training-free framework for physics-aware fluid-object interaction video generation. The core idea is to keep pretrained video models doing what they already do well—appearance synthesis—while handing motion dynamics to a physics simulator. That separation matters for teams experimenting with generative video …

• Fluid-Gen-Zero is a training-free framework for physics-aware fluid-object interaction video generation.
• Hong Huang and five coauthors; published on arXiv cs.GR / cs.CV.

arXiv cs.GR details BrickBench, a benchmark for agentic text-conditioned LEGO-set design that asks an AI agent to produce an assembly matching a prompt while still being physically buildable. The setup pushes models to choose from a discrete parts library and reason across both local fit …

• BrickBench benchmarks agentic text-conditioned LEGO-set design against physical and semantic constraints.
• Peter Kulits, Yiqing Xu, R. Kenny Jones, Cordelia Schmid, and Jiajun Wu.