3D Aesthetics is significant in digital design, shaping how users experience real-time 3D content in games, VR, and product design. However, creating aesthetically pleasing shapes remains challenging due to diverse subjective standards and the lack of tools that support aesthetics-driven editing. Users often rely on intuition without explicit guidance on visual appeal, making aesthetics refinement slow, inconsistent, and cognitively demanding, particularly in fast-paced, iterative workflows. To address this challenge, we conducted in-depth interviews with design experts to identify challenges in aesthetics-oriented modeling workflows. Based on the findings, we developed Aesthetic3D , a 3D modeling interface that provides real-time aesthetics scores learned from human perceptual data. Furthermore, Aesthetic3D seamlessly integrates the learned aesthetics measures into intuitive editing operations, enabling aesthetics-driven exploration and refinement of shape geometry. We evaluated Aesthetic3D through an ablation study, an open-ended study, and three generalization evaluations. Comprehensive experiments show that with Aesthetic3D , users can easily and effectively enhance the aesthetics appeal of 3D shapes.
Implicit surfaces offer distinct advantages over traditional boundary representations, including infinite resolution, low memory footprint, smooth geometry by construction, and support for non-destructive modeling. In this work, we introduce a method for localizing geometric detail in a way that preserves the mathematical properties required for accurate and efficient rendering using sphere tracing. Our contributions include novel procedural modeling techniques that expand the range of repetition patterns achievable in implicit surfaces; an interpolation-based approach that maintains field correctness while remaining computationally efficient; and a cache-based acceleration strategy that significantly improves the rendering performance of domain-repeated implicit geometries.
Welcome to the 2026 I3D PACMCGIT issue. PACMCGIT is a specialized journal in the collection of Proceedings of the ACM (PACM) that is specifically focused on publishing research of the highest quality within the broad domains of computer graphics and interactive techniques. This issue is a collection of the papers selected and presented at the 30th ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games (I3D) 2026. Continuing the tradition, the I3D collection of papers constitutes the first PACMCGIT issue of the year, this time marking the beginning of PACMCGIT’s ninth year presenting the publications of I3D. I3D 2026 received 41 submissions. Following a rigorous double-blind peer-reviewing process with 122 reviews, including a second review cycle by a panel of experts from the I3D 2026 international program committee, 16 papers were accepted to be presented at the conference, resulting in an acceptance rate of \(39\%\) . This year, one paper was desk-rejected as it was clearly incomplete. Paper topics span various exciting areas, from physical simulation and animation to real-time rendering, ray-tracing, VR, and neural rendering. These papers appeared in talk presentations at the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games (I3D) in May 2026 at Lucasfilm in San Francisco (CA).
The generation of consistent multi-view images is essential for diffusion-based 3D texture synthesis, a key component of modern 3D content pipelines. However, existing methods remain weakly coupled across viewpoints, often exhibiting inconsistencies between views that result in artifacts such as ghosting, misalignment, and texture distortions. We introduce PixelSync, a training-free mechanism that explicitly enforces multi-view consistency during the diffusion process. Our method operates in two stages. First, we compute visibility-aware pixel correspondences between known views in a fast pre-computation step by utilizing geometric information obtained from a given 3D model. Then, during the diffusion process, we enforce consistency through two complementary algorithms: (i) attention synchronization, by modifying the attention scores in the U-Net’s down- and mid-blocks, and (ii) latent synchronization, by progressively aligning features on their corresponding shared pixels. PixelSync can be seamlessly integrated into pre-trained diffusion models without retraining, offering a lightweight and effective solution. We demonstrate the effectiveness of our approach for generative texture synthesis by plugging PixelSync into established multi-view diffusion backbones, showcasing reduced cross-view artifacts and improved global coherence across standard quantitative and qualitative evaluations.
This paper addresses the problem of unpaired shape translation on 3D point clouds. While prior methods typically rely on global latent vectors or spatially structured grids, such representations often lack the flexibility to capture both semantic structures and fine-grained geometric details. To address this, we propose operating directly in a structured token space, where tokens are pretrained through masked autoencoding. Unlike rigid spatial grids that impose fixed layouts, our tokens naturally adapt to geometric variations while maintaining semantic coherence. This structured yet flexible latent space enables semantically meaningful and geometrically precise transformations. A transformer-based translator is proposed to manipulate these tokens. This gated dual-branch translator enables detail-preserving and topology-aware shape translation across categories. Experiments on challenging tasks, such as chair-to-table transformations, demonstrate that our approach outperforms existing methods in preserving both global structure and part-level details.
We introduce Canvas3D, a novel framework for learning an editable 3D instance field from sparse 2D observations. Our approach utilizes depth information from observed views to warp sparse inputs into previously unobserved viewpoints. Each missing view is warped from the two nearest observed views in different directions, effectively filling occlusion-caused gaps. These warped views significantly enhance the learning of a geometrically precise 3D instance field by leveraging multi-view consistency. Additionally, we propose an effective method for manipulating the learned 3D instance field through 2D interactions in a geometrically coherent manner. This allows for object-level manipulation in 3D space by editing 2D instance maps from arbitrary viewpoints. The instance field also serves as an intermediate representation for 3D-aware editable image synthesis. This bypasses the complexities of direct 3D manipulation by editing a robust instance field before translating it into consistent images using off-the-shelf semantic image synthesis models. Extensive experiments demonstrate that our method accurately reconstructs a 3D instance field from sparse 2D observations and enables 2D-driven object-level manipulation of the field. We also explore the potential of our method in 3D-aware controllable image synthesis, showcasing synthesis results guided by edited semantic maps.
Algorithms leveraging ReSTIR-style spatiotemporal reuse have recently proliferated, hugely increasing effective sample count for light transport in real-time ray and path tracers. Many papers have explored novel theoretical improvements, but algorithmic improvements and engineering insights toward optimal implementation have largely been neglected. We demonstrate enhancements to ReSTIR PT that make it 2-3 & times; faster, decrease both visual and numerical error, and improve its robustness, making it closer to production-ready. We halve the spatial reuse cost by reciprocal neighbor selection, robustify shift mappings with new footprint-based reconnection criteria, and reduce spatiotemporal correlation with duplication maps. We further improve both performance and quality by extensive optimization, unifying direct and global illumination into the same reservoirs, and utilizing existing techniques for color noise and disocclusion noise reduction.
Frametime spikes and graphical fidelity both matter for the feel of first-person shooter (FPS) games, yet their combined effects are not well understood. This paper examines how graphics settings and frametime spikes during aiming interact with player performance and experience. We developed a custom FPS game with configurable textures, lighting, and visual effects, and induced frametime spikes of 0 ms, 225 ms, or 675 ms during play. Twenty-one participants completed all combinations of graphics and spikes while providing performance data and subjective ratings for visual quality and smoothness. Results show that graphics quality primarily affects perceived visual quality , while frametime spikes primarily reduce perceived smoothness . Performance (i.e., score and accuracy) declines with larger spikes but remains largely unchanged across graphics settings. These findings suggest that spike magnitude, rather than graphics quality, is the dominant factor shaping smoothness and performance, while players still notice higher-quality graphics.
High-quality facial animation retargeting is crucial in digital entertainment production, enabling the transfer of animations from a source rig to a target rig. To avoid the need of manual transfer, automatic retargeting methods can be employed. However, creating universal methods remains challenging due to varying rig and topologies. We propose a novel retargeting method based on the idea that any facial expression can be decomposed into a weighted sum of Facial Action Coding System (FACS) poses. Our method assumes that these decompositions transfer across rigs with minimal need of refinement. We utilize a machine learning model to extract FACS weights from source frames. This model is trained solely on source animation data, requiring no target data or complex correspondence mappings. After a refinement step, these weights are then applied to the target rig’s FACS poses. Our technique supports seamless animation transfer across different rigs while allowing for artistic modifications by adjusting individual FACS poses. Our experiments demonstrate that this method produces high-quality retargeted animations and outperforms other retargeting methods. We explore how different design choices impact retargeting quality and showcase artistic control of asymmetries. By simplifying the workflow for rigged heads, this flexible framework enhances production efficiency while maintaining artistic control.
Virtual Reality offers highly immersive experiences, yet users remain physically situated in real-world environments whose layouts often differ from the virtual spaces they perceive. In these settings, users’ spatial memory of the physical environment may interact with or even conflict with the virtual scene, potentially influencing locomotion, task performance, and situational awareness. Despite its importance, the influence of physical space memory on VR interaction remains largely unexplored. To bridge this gap, we conducted three controlled studies to examine how physical-space memory influences user behavior in VR across different conditions. Specifically, we investigated how the proximity of remembered physical obstacles shapes users’ immediate locomotion and interaction strategies, how reliance on physical-space memory evolves over time as virtual knowledge accumulates during continuous tasks, and how varying levels of spatial consistency between virtual layouts and remembered physical landmarks affect user behavior. Our findings reveal that users subconsciously adapt their navigation and decision-making based on remembered surroundings, even when no visual cues are available. These insights deepen our understanding of spatial cognition in immersive VR and provide actionable guidelines for designing safer, more efficient, and cognitively consistent virtual environments.
To enhance learning and training, VR haptic devices must provide safe and physically accurate feedback that helps users build intuitions about force and motion. We present TrueForce, a method for rendering force feedback by capturing user input and reconfiguring a compliant mechanism in real time to align perceived with expected forces. We implemented TrueForce in a grounded Encountered-Type Haptic Display (ETHD) and demonstrated its versatility through three representative physics scenarios: pushing a stationary object, experiencing electrostatic repulsion, and bouncing a moving object. A user study with 29 participants showed that users adapted to new interaction profiles within 2-3 trials, and that rendered forces remained within perceptual thresholds except in the heaviest and most dynamic conditions. These results suggest that users can rapidly adapt to the system's force-motion mappings, providing the consistent sensorimotor feedback necessary for building physical intuitions when interacting with moving virtual objects.
Collaborative tasks in Virtual Reality, such as jointly manipulating objects, require tight multi-user synchronization. However, users often operate in heterogeneous physical environments with varying room sizes, furniture layouts, and obstacles. Enforcing a single, globally consistent virtual scene across such diversity constrains users’ natural movements, underutilizes available physical space, and ultimately degrades collaboration efficiency. We present CoVR, a novel framework that enables seamless multi-user collaboration in VR across heterogeneous physical spaces by prioritizing interaction alignment over strict spatial alignment. Instead of requiring all users to share an identical virtual scene, CoVR provides each user with a personalized virtual scene layout while maintaining synchronized actions and consistent collaborative outcomes. Technically, we formulate a multi-user coupled optimization problem that jointly optimizes per-user layouts to maximize usable space and facilitate seamless task execution. A key challenge arises from bi-directional dependencies: one user’s optimized layout determines feasible trajectories, which in turn constrain other users’ optimizations, and vice versa. To tackle this, CoVR introduces an iterative joint optimization algorithm alternating between personalized scene refinement and cross-user synchronization until convergence. We evaluate CoVR across typical collaborative tasks. Results show that CoVR significantly outperforms conventional scene alignment methods, leading to notable improvements in spatial efficiency, task completion time, and user comfort. These findings underscore its potential for scalable, natural, and efficient multi-user VR collaboration.
Virtual reality (VR) applications face a critical trade-off: while visual immersion is the core experience, it often induces cybersickness (CS), leading to user attrition. Existing visual interventions (e.g., field-of-view restriction) alleviate discomfort at the cost of compromising immersion. Auditory interventions have been widely applied in mitigating motion sickness (MS), such as through music interventions and synchronized sound effects. Among all MS mitigation approaches, anticipatory audio cues have demonstrated effectiveness in vehicular MS contexts, reducing symptoms by approximately 17%. However, questions remain regarding VR cybersickness (CS): Do anticipatory audio cues remain effective in highly immersive virtual environments, where the dominance of visual interaction in VR may alter cue perception compared to that in MS contexts? How can VR motion intensity information be encoded, given that such information is largely absent in current mitigation strategies? How can anticipatory audio cues be systematically designed for CS mitigation through explicit mapping frameworks that link visual motion parameters to audio features? To address these gaps, this study establishes an optical flow mapping framework that encodes motion direction through stereo positioning and intensity through temporal patterns, delivering the results 2 seconds in advance. A within-subjects experiment (N = 24) was conducted in VR roller coaster environments, in which four conditions were compared. Results demonstrated a significant reduction in cybersickness across all audio conditions (direction-only: d = 0.862; intensity-only: d = 0.670; combined: d = 1.497, p < .001), with combined cues outperforming single-dimension alternatives (p < .016). All interventions preserved visual immersion (M approximate to 4.0/5) and maintained high usability metrics. This framework demonstrates that non-visual predictive cues can address the immersion-comfort trade-off in VR applications.
3D Gaussian Splatting (3DGS) is a popular representation for radiance field reconstruction, distinguished by the rendering speed of its rasterization-based renderer. While 3D Gaussians can also be ray traced, this approach has so far been slower, with 3D Gaussian Ray Tracing (3DGRT) taking nearly one order of magnitude longer to optimize. To address this, we present GRay, a fast ray tracer for 3D Gaussians designed to close this performance gap and match 3DGS's speed. Our method leverages the algorithmic difference between both approaches: unlike rasterization, ray tracing evaluates only Gaussians that are actually intersected by a ray, leading to potentially logarithmic–rather than linear–scaling in the number of primitives. This property allows ray tracing to better exploit dense scenes composed of numerous tiny Gaussians, a configuration which has largely been overlooked. Notably, we show that dense initialization–which creates many small Gaussians–slows down rasterization, but instead speeds up ray tracing. Designed to leverage this effect, GRay renders nearly 4x faster and optimizes nearly 10x faster than 3DGRT while maintaining similar quality, and has competitive speed with 3DGS albeit at somewhat lower quality. Code is available at https://repo-sam.inria.fr/nerphys/gray.
Intrinsic decomposition which expresses image colors as the product of diffuse albedo and shading, possibly augmented with view-dependent residuals has a long history in image editing as it enables the modification of object colors and textures without altering lighting. We extend intrinsic decomposition to radiance fields represented with Gaussian splatting by proposing solutions to three key aspects of such decomposition. First, we describe how to model the intrinsic decomposition as independent sets of Gaussian primitives, which allows each set to adapt to the characteristics of the layer it represents. Second, we present an optimization procedure guided by data-driven predictions to disentangle multi-view photographs of a scene into the aforementioned intrinsic sets. Finally, we provide an editing workflow where users modify the texture of planar surfaces simply by modifying the albedo of that surface in one image. Capturing this edit within the intrinsic radiance field allows re-rendering of the edited scene with plausible lighting under arbitrary viewpoints.
Abstract We present a technique for simulating detailed cloth dynamics on coarse meshes at interactive rates. Building on the wrinkle parameterization introduced by [CCK*21], we introduce a dynamics model for a cloth sheet represented as a superposition of a base surface and a wrinkle distribution parameterized by spatially varying amplitude and frequency. Our model incorporates two‐way coupling of the base surface and the wrinkle parameters, allowing the base deformation to drive emergence of wrinkles and permitting compression of the base surface in turn. To deform the mesh using the computed wrinkle parameters, we also introduce a simple phase reconstruction strategy that produces dynamically evolving, temporally coherent wrinkles on the simulated mesh.
