These 41 SIGGRAPH Asia 2025 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every SIGGRAPH Asia 2025 paper, free trial on arXivSub.
3D SMoE Splatting for Edge-aware Realtime Radiance Field Rendering
Yi-Hsin Li, Mårten Sjöström
CodeOptimizationComputational EfficiencyMixture of ExpertsNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: Proposed a new 3D SMoE Splatting (3DSMoES) framework, integrating the Steered Mixtures-of-Experts regression method into the CUDA-optimized process of 3D Gaussian Splatting (3DGS) to achieve 3D rendering.
A Stack-Free Parallel h-Adaptation Algorithm for Dynamically Balanced Trees on GPUs
Lixin Ren, Enhua Wu
CodeOptimizationComputational Efficiency
🎯 What it does: Developed a stackless parallel h-adaptive algorithm for GPUs for the construction and update of dynamic balanced trees to improve fluid simulation performance
AutoBrep : Autoregressive B-Rep Generation with Unified Topology and Geometry
Xiang Xu (Autodesk Research), Peter Meltzer
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: Proposed and trained a single autoregressive Transformer called AutoBrep, which generates complete B-Reps using a unified discrete token sequence based on face-edge topology, supporting auto-completion.
🎯 What it does: Propose a complete display evaluation system called CameraVDP, which combines a calibrated consumer-grade camera with a visual difference predictor. It can achieve precise photometric, color, and geometric correction on the original images captured by the camera, and assess human-perceptible display defects through the visual difference predictor, while providing uncertainty estimation.
🎯 What it does: Propose an inverse operation (accretion and erosion) planning algorithm that can generate executable hybrid additive-subtractive manufacturing sequences for any complex geometry, ensuring that the final part is completely consistent with the target geometry.
🎯 What it does: An improved method based on clustering error correction and group dynamic transformation is proposed for 4D Gaussian point cloud rendering. It uses elliptical error clustering to locate missing or occluded errors and dynamically adds new Gaussian points through back-projection or foreground segmentation. At the same time, group-shared temporal transformations are introduced to maintain consistent correspondence between dynamic objects and point clouds.
🎯 What it does: Propose DvD, a document dewarping method based on coordinate-level diffusion models, which generates dewarping mappings instead of directly generating images, and construct a large-scale multi-domain evaluation benchmark called AnyPhotoDoc6300.
🎯 What it does: This paper proposes a 'FreeMusco' framework that does not rely at all on motion capture data. It can jointly learn the latent gait representation and low-level control strategies of musculoskeletal characters, achieving energy-aware, morphology-adaptive diverse gaits, and can further be used for high-level tasks such as target navigation and path following.
🎯 What it does: Proposes GigaSLAM, a monocular RGB odometry and mapping framework that utilizes a hierarchical sparse voxel structure and Gaussian projection rendering, enabling real-time localization and high-quality scene rendering in kilometer-scale, unbounded outdoor environments.
High-Fidelity Dynamic Portrait Animation via Direct Preference Optimization and Temporal Motion Modulation
Jiahao Cui, Siyu Zhu
CodeGenerationData SynthesisReinforcement Learning from Human FeedbackConvolutional Neural NetworkTransformerDiffusion modelImageVideoAudio
🎯 What it does: Proposes a diffusion framework based on human preference alignment for high-dynamic, realistic portrait animation driven by audio and skeletal motion.
🎯 What it does: Proposed an end-to-end single-phone scanning workflow that utilizes two-phase static and dynamic video sequences to construct high-fidelity full-body clothing avatars that can be rendered in real-time on mobile devices;
Imaginarium: Vision-guided High-Quality 3D Scene Layout Generation
Xiaoming Zhu (Tsinghua University), Long Zeng (Tsinghua University)
CodeGenerationPose EstimationDepth EstimationRetrievalTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageTextPoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: A visually driven system is proposed, which can generate high-quality, logically coherent 3D scene layouts from a predefined 3D asset library by combining text prompts, refined image generation models, image parsing, and asset retrieval.
🎯 What it does: A two-stage framework called Img2CAD is constructed, which first uses a VLM to predict the discrete CAD structure from a single-view image, and then employs a Transformer to predict continuous attributes, thereby achieving reverse engineering from images to 3D CAD.
🎯 What it does: Propose a conditional video distillation method that realizes real-time lip-sync video generation using input human pose information, combining input-aware sparse attention and distillation loss.
Kaixuan Wei (King Abdullah University of Science and Technology), Wolfgang Heidrich (King Abdullah University of Science and Technology)
CodeOptimizationConvolutional Neural NetworkDiffusion modelAuto EncoderImagePoint CloudTabularPhysics Related
🎯 What it does: We have designed and implemented a differentiable optical design framework aimed at large-area, mass-producible applications, taking into account process distortions and numerical high-resolution capabilities. It combines grayscale direct-write photolithography plus nanoimprint replication, utilizing a neural network digital twin to achieve process modeling, and employs GSPMD distributed FFT and convolution for large-scale DOE optimization.
Giancarlo Pereira (New York University), D. Panozzo
CodeDepth EstimationGaussian SplattingSimultaneous Localization and MappingOptical FlowPoint CloudMesh
🎯 What it does: Propose LookUp3D, a structured light 3D scanning method based on a per-pixel color-to-depth lookup table, capable of acquiring high-resolution 3D geometry at high frame rates.
🎯 What it does: Proposes a parameter-efficient update method for instance-specific neural fields using low-rank adaptation (LoRA), which can encode minor edits (such as image filtering, geometric deformation, video compression, energy minimization) as low-rank weight increments;
🎯 What it does: A fully unlabeled voice-driven 3D facial animation framework, LSF-Animation, was developed, which can directly extract emotional and identity features from raw audio and neutral facial meshes, achieving unlabeled generation.
🎯 What it does: Proposed a multi-frame monocular depth estimation system called MODepth, which is based on controlling motion through an optical image stabilization (OIS) module, and designed the MODNet network as well as the principal point offset estimation module and the pose estimation module.
🎯 What it does: By introducing the Neural Octahedral Field (NOF) as a prior, simultaneous point cloud denoising and adaptive regularization of smooth and sharp edges are achieved for neural implicit surfaces, thereby recovering clearer and more sharply featured 3D surfaces from noisy point clouds.
🎯 What it does: This paper proposes a deep learning-based method for point cloud visibility determination, directly predicting the visibility of each point under a given viewpoint;
🎯 What it does: Propose NeuralPVS, a real-time depth learning-based method for computing potential visibility sets (PVS) from regions, capable of outputting nearly error-free visible geometry sets at a frame rate of 100 Hz.
🎯 What it does: Propose OmnimatteZero, a fully training-free real-time video matting method that achieves object removal, foreground extraction, and layer composition while preserving object effects such as shadows and reflections.
🎯 What it does: This paper proposes a novel graph learning framework called PoissonNet, which combines local gradient transformation and global Poisson equation solving to achieve efficient learning of surface features on triangular meshes.
🎯 What it does: Proposes the PowerGS framework, which jointly optimizes the rendering power consumption and display power consumption of the 3D Gaussian Splatting (3DGS) model under the constraint of visual quality, and supports disparity rendering.
🎯 What it does: This paper proposes PractiLight, a method for controllable relighting that utilizes the base diffusion model (Stable Diffusion 1.5) to manipulate lighting and shadows. It trains a lightweight LoRA regressor on the self-attention layer to predict the direct-irradiance map of an image, and during image generation, it employs Classifier Guidance, ControlNet, and self-attention query injection techniques to freely modify the light source position, intensity, and material on the generated image.
🎯 What it does: Propose a single-frame structured light 3D reconstruction framework called NSL based on neural feature decoding, which can achieve high-precision depth recovery under a single camera + projection mode;
🎯 What it does: Proposes a self-supervised underwater color restoration framework based on waveform diffusion models, utilizing filtered multi-scale feature distillation;
🎯 What it does: Proposes a shape-aware sparse inertial capture system, SAIP, which can achieve real-time motion capture for different body types (including children) using only six IMUs, and first realizes human shape estimation in IMU-based capture.
🎯 What it does: Propose the Sorted Opacity Fields (SOF) method, enabling the fast extraction of high-quality unbounded surface meshes from 3D Gaussian representations.
🎯 What it does: Propose STGlight, a lightweight online indoor lighting estimation method that can process LDR RGB-D video streams, maintain incrementally updated global geometric and lighting representations, and predict HDR environment maps for any position in each frame.
🎯 What it does: Proposes the Teamwork framework, which utilizes multi-instance low-rank adaptation and coordination mechanisms to achieve input/output channel expansion and task transfer without modifying the structure of pre-trained diffusion models.
Temporally Smooth Mesh Extraction for Procedural Scenes with Long-Range Camera Trajectories using Spacetime Octrees
Zeyu Ma (Princeton University), Jia Deng (Princeton University)
CodeGenerationData SynthesisNeural Radiance FieldGaussian SplattingOptical FlowPoint CloudMeshTime Series
🎯 What it does: Propose the BinocMesher method, which extracts temporally smooth 3D meshes from an occupancy function under long-distance camera trajectories using a 4D Binary-Octree.
🎯 What it does: By combining pixel trajectories from multi-view point tracking with a 3D Gaussian model, TrackerSplat rapidly and robustly reconstructs dynamic 3D scenes in multi-view videos.
🎯 What it does: Propose the Uni3C framework, unifying 3D-enhanced camera trajectory control with human motion control, for generating high-quality controllable videos from single views.
UTDesign: A Unified Framework for Stylized Text Editing and Generation in Graphic Design Images
Yiming Zhao (Peking University), Zhouhui Lian (Peking University)
CodeImage TranslationGenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelDiffusion modelRectified FlowAuto EncoderImageTextMultimodalityBenchmark
🎯 What it does: Propose the UTDesign unified framework to achieve high-precision stylized text editing and generation of English labels in design images, and integrate a complete text-to-design (T2D) pipeline.
🎯 What it does: This paper proposes an encoding method that directly stores learnable features on the vertices of a 3D mesh (Neural Vertex Features), and combines adaptive multi-resolution refinement (LOD) to achieve efficient neural global illumination rendering.
🎯 What it does: Propose a computational framework that strictly considers global and local collision constraints to plan the addition and subtraction of material hybrid remanufacturing processes in order to maximize the reusable volume.