Computer Graphics Forum最新文献

The utilization of neural implicit surface as a geometry representation has proven to be an effective multi-view surface reconstruction method. Despite the promising results achieved, reconstructing geometry from objects in real-world scenes remains challenging due to the interaction between surface materials and complex ambient light, as well as shadow effects caused by self-occlusion, making it a highly ill-posed problem. To address this challenge, we propose MF-SDF, a method that use a hybrid neural network and spherical gaussian representation to model environmental lighting, so that the model can express the situation of multiple light sources including directional light (such as outdoor sunlight) in real-world scenarios. Benefit from this, our method effectively reconstructs coherent surfaces and accurately locates the shadow location on the surface. Furthermore, we adopt a shadow aware multi-view photometric consistency loss, which mitigates the erroneous reconstruction results of previous methods on surfaces containing shadows, thereby improve the overall smoothness of the surface. Additionally, unlike previous approaches that directly optimize spatial features, we propose a Fourier feature optimization method that directly optimizes the tensorial feature in the frequency domain. By optimizing the high-frequency components, this approach further enhances the details of surface reconstruction. Finally, through experiments, we demonstrate that our method outperforms existing methods in terms of reconstruction accuracy on real captured data.

Swept volume computation—the determination of regions occupied by moving objects—is essential in graphics, robotics, and manufacturing. Existing approaches either explicitly track surfaces, suffering from robustness issues under complex interactions, or employ implicit representations that trade off geometric fidelity and face optimization difficulties. We propose a novel inversion of motion perspective: rather than tracking object motion, we fix the object and trace spatial points backward in time, reducing complex trajectories to efficiently linearizable point motions. Based on this, we introduce a multi-field tetrahedral framework that maintains multiple distance fileds per element, preserving fine geometric details at trajectory intersections where single-field methods fail. Our method robustly computes swept volumes for diverse motions, including translations and screw motions, and enables practical applications in path planning and collision detection.

Although recent text-to-image (T2I) diffusion models excel at aligning generated images with textual prompts, controlling the visual style of the output remains a challenging task. In this work, we propose Style-Prompting Guidance (SPG), a novel sampling strategy for style-specific image generation. SPG constructs a style noise vector and leverages its directional deviation from unconditional noise to guide the diffusion process toward the target style distribution. By integrating SPG with Classifier-Free Guidance (CFG), our method achieves both semantic fidelity and style consistency. SPG is simple, robust, and compatible with controllable frameworks like ControlNet and IPAdapter, making it practical and widely applicable. Extensive experiments demonstrate the effectiveness and generality of our approach compared to state-of-the-art methods. Code is available at https://github.com/Rumbling281441/SPG.

Dynamic scene rendering has taken a leap forward with the rise of 4D Gaussian Splatting, but there is still one elusive challenge: how to make 3D Gaussians move through time as naturally as they would in the real world, all while keeping the motion smooth and consistent. In this paper, we present an approach that blends state-space modeling with Wasserstein geometry, enabling a more fluid and coherent representation of dynamic scenes. We introduce a State Consistency Filter that merges prior predictions with the current observations, enabling Gaussians to maintain coherent trajectories over time. We also employ Wasserstein Consistency Constraint to ensure smooth, consistent updates of Gaussian parameters, reducing motion artifacts. Lastly, we leverage Wasserstein geometry to capture both translational motion and shape deformations, creating a more geometrically consistent model for dynamic scenes. Our approach models the evolution of Gaussians along geodesics on the manifold of Gaussian distributions, achieving smoother, more realistic motion and stronger temporal coherence. Experimental results show consistent improvements in rendering quality and efficiency.

(see https://www.acm.org/publications/class-2012)

Recent large reconstruction models have made notable progress in generating high-quality 3D objects from single images. However, current reconstruction methods often rely on explicit camera pose estimation or fixed viewpoints, restricting their flexibility and practical applicability. We reformulate 3D reconstruction as image-to-image translation and introduce the Relative Coordinate Map (RCM), which aligns multiple unposed images to a “main” view without pose estimation. While RCM simplifies the process, its lack of global 3D supervision can yield noisy outputs. To address this, we propose Relative Coordinate Gaussians (RCG) as an extension to RCM, which treats each pixel's coordinates as a Gaussian center and employs differentiable rasterization for consistent geometry and pose recovery. Our LucidFusion framework handles an arbitrary number of unposed inputs, producing robust 3D reconstructions within seconds and paving the way for more flexible, pose-free 3D pipelines.

Point cloud normal estimation underpins many 3D vision and graphics applications. Precise normal estimation in regions of sharp curvature and high-frequency variation remains a major bottleneck; existing learning-based methods still struggle to isolate fine geometry details under noise and uneven sampling. We present FAHNet, a novel frequency-aware hierarchical network that precisely tackles those challenges. Our Frequency-Aware Hierarchical Geometry (FAHG) feature extraction module selectively amplifies and merges cross-scale cues, ensuring that both fine-grained local features and sharp structures are faithfully represented. Crucially, a dedicated Frequency-Aware geometry enhancement (FA) branch intensifies sensitivity to abrupt normal transitions and sharp features, preventing the common over-smoothing limitation. Extensive experiments on synthetic benchmarks (PCPNet, FamousShape) and real-world scans (SceneNN) demonstrate that FAHNet outperforms state-of-the-art approaches in normal estimation accuracy. Ablation studies further quantify the contribution of each component, and downstream surface reconstruction results validate the practical impact of our design.

Although single-image 3D human reconstruction has made significant progress in recent years, few of the current state-of-the-art methods can accurately restore the appearance and geometric details of loose clothing. To achieve high-quality reconstruction of a human body wearing loose clothing, we propose a learnable dynamic adjustment framework that integrates side-view features and the uncertainty of the parametric human body model to adaptively regulate its reliability based on the clothing type. Specifically, we first adopt the Vision Transformer model as an encoder to capture the image features of the input image, and then employ SMPL-X to decouple the side-view body features. Secondly, to reduce the limitations imposed by the regularization of the parametric model, particularly for loose garments, we introduce a learnable coefficient to reduce the reliance on SMPL-X. This strategy effectively accommodates the large deformations caused by loose clothing, thereby accurately expressing the posture and clothing in the image. To evaluate the effectiveness, we validate our method on the public CLOTH4D and Cape datasets, and the experimental results demonstrate better performance compared to existing approaches. The code is available at https://github.com/yyd0613/CoRe-Human.

Using 3D Gaussian splatting to reconstruct large-scale aerial scenes from ultra-high-resolution images is still a challenge problem because of two memory bottlenecks - excessive Gaussian primitives and the tensor sizes for ultra-high-resolution images. In this paper, we propose a task partitioning algorithm that operates in both object and image space to generate a set of small-scale subtasks. Each subtask's memory footprints is strictly limited, enabling training on a single high-end consumer-grade GPU. More specifically, Gaussian primitives are clustered into blocks in object space, and the input images are partitioned into sub-images according to the projected footprints of these blocks. This dual-space partitioning significantly reduces training memory requirements. During subtask training, we propose a depth comparison method to generate a mask map for each sub-image. This mask map isolates pixels primarily contributed by the Gaussian primitives of the current subtask, excluding all other pixels from training. Experimental results demonstrate that our method successfully achieves large-scale aerial scene reconstruction using 9K resolution images on a single RTX 4090 GPU. The novel views synthesized by our method retain significantly more details than those from current state-of-the-art methods.

Simulating wet clothing remains a significant challenge due to the complex physical interactions between moist fabric and the human body, compounded by the lack of dedicated datasets for training data-driven models. Existing self-supervised approaches struggle to capture moisture-induced dynamics such as skin adhesion, anisotropic surface resistance, and non-linear wrinkling, leading to limited accuracy and efficiency. To address this, we present SHGS, a novel self-supervised framework for humidity-controllable clothing simulation grounded in the physical modeling of capillary bridges that form between fabric and skin. We abstract the forces induced by wetness into two physically motivated components: a normal adhesive force derived from Laplace pressure and a tangential shear-resistance force that opposes relative motion along the fabric surface. By formulating these forces as potential energy for conservative effects and as mechanical work for non-conservative effects, we construct a physics-consistent wetness loss. This enables self-supervised training without requiring labeled data of wet clothing. Our humidity-sensitive dynamics are driven by a multi-layer graph neural network, which facilitates a smooth and physically realistic transition between different moisture levels. This architecture decouples the garment's dynamics in wet and dry states through a local weight interpolation mechanism, adjusting the fabric's behavior in response to varying humidity conditions. Experiments demonstrate that SHGS outperforms existing methods in both visual fidelity and computational efficiency, marking a significant advancement in realistic wet-cloth simulation.

Neural signed-distance fields (SDFs) are a versatile backbone for neural geometry representation, but enforcing CAD-style developability usually requires Gaussian-curvature penalties with full Hessian evaluation and second-order differentiation, which are costly in memory and time. We introduce an off-diagonal Weingarten loss that regularizes only the mixed shape operator term that represents the gap between principal curvatures and flattens the surface. We present two variants: a finite-difference version using six SDF evaluations plus one gradient, and an auto-diff version using a single Hessian-vector product. Both converge to the exact mixed term and preserve the intended geometric properties without assembling the full Hessian. On the ABC benchmarks the losses match or exceed Hessian-based baselines while cutting GPU memory and training time by roughly a factor of two. The method is drop-in and framework-agnostic, enabling scalable curvature-aware SDF learning for engineering-grade shape reconstruction. Our code is available at https://flatcad.github.io/.