100k CAD models released as executable Python programs
Each sample pairs its construction code with validated feature history, STEP geometry, and point clouds after kernel checks.
cs.GR
Graphics
Covers all aspects of computer graphics. Roughly includes material in all of ACM Subject Class I.3, except that I.3.5 is is likely to have Computational Geometry as the primary subject area.
sort pith recommended most recent
One image yields a mobile-ready 3D asset in 30 seconds
AssetGen controls polygon count while baking normals and textures so the output runs in real time without further editing.
· “AssetGen: Deployable 3D Asset Generation at Interactive Speed”
Polynomial kernel sharpens 3D surfaces from multi-view photos
The local-support kernel replaces exponential smoothing in implicit moving least squares to retain fine geometry and enable direct end-toend
· “High-Fidelity Surface Splatting-Based 3D Reconstruction from Multi-View Images”
Semantic codebook creates style-matched co-speech gestures
By organizing motion codes according to gesture semantics and applying reference prompts, the system produces motions faithful to both word,
Algebraic coarsening delivers 3x speedup in GPU contact solves
In-solve vertex aggregation reduces system size without remeshing or loss of contact accuracy.
· “AGIPC: Adaptive In-Solve Algebraic Coarsening for GPU IPC”
Cumulative maps let diffusion models generate in one step
By tying short updates to full transport, existing models reach high-quality samples faster with minimal changes to training.
Retrieval lifts Blender code success from 41% to 70%
A fixed set of 500 examples makes existing LLMs write reliable 3D modeling scripts on ordinary hardware.
· “BlenderRAG: High-Fidelity 3D Object Generation via Retrieval-Augmented Code Synthesis”
Taxonomy shows most 3D generators miss engine-ready assets
Organizing methods by asset tiers and lifecycle stages exposes the gap between visual synthesis and production standards for games, AI, and
· “From Visual Synthesis to Interactive Worlds: Toward Production-Ready 3D Asset Generation”
Builds an animatable 3D human in one forward pass
A canonical Gaussian asset predicted from sparse views is re-posed via skinning, cutting per-frame synthesis to 0.037 s.
· “Real-Time Human Reconstruction and Animation using Feed-Forward Gaussian Splatting”
Render gradients on deformation gradient reduce morphing silhouette error
Injecting visual supervision via F in MPM with Gaussian splatting yields up to 50% lower silhouette error on thin features.
· “PhysMorph-GS: Render-Guided Volumetric Morphing with Differentiable Physics”
Dual-path diffusion sharpens lip sync and head poses
Separating expression and pose features while highlighting intense motion frames produces more accurate talking-head video from audio.
· “KSDiff: Keyframe-Augmented Speech-Aware Dual-Path Diffusion for Facial Animation”
Exact mass preservation comes to diffusion models
Each intensity unit is a hopping particle, so rock porosity and battery phase fractions come out exactly right.
· “Discrete Spatial Diffusion: Intensity-Preserving Diffusion Modeling”
3D scene graph plans portraits before shutter click
Generates human pose, camera, and lighting setups that raters prefer over post-capture baselines while staying physically valid.
· “Before the Shutter: Aesthetic and Actionable Portrait Photography Planning in 3D Scenes”
One diffusion checkpoint stipples any density at fixed cost
Late-stage ControlNet conditioning plus gated projection lets a single model match optimized baselines while generalizing to unseen pointbud
· “StippleDiffusion: Capacity-Constrained Stippling using Controlled Diffusion”
WildRelight dataset adapts synthetic relighting models to real scenes on the fly
Temporal sequences of natural light supply self-supervision that aligns models with outdoor statistics without ground-truth relit images.
· “WildRelight: A Real-World Benchmark and Physics-Guided Adaptation for Single-Image Relighting”
Precomputed maps simulate real lens optics an order of magnitude faster
A factorized model with wavelength inputs and a binary ray mask reproduces flares and aberrations without per-wavelength polynomials or full
NeRF-based capture wins for lab-object holograms
Seventeen graduate raters scored 48 holograms; the neural method led on reflective and low-texture labware.
A 9-point curve scheme gains three shape-control knobs
Blending two 7-point refinements yields smooth approximating curves with local shape control.
· “A Three-Parameter Binary Subdivision Scheme for Shape-Controlled Curve Design”
CVAE teaches virtual faces emotions from just 7,680 samples
A small generative model plus residual refinement produces intensity-controlled expressions without actor capture.
· “Emotion Intensity Matters: Generating Realistic Expressions in Virtual Humans with CVAEs”
A 3D generator that obeys part names and bounding boxes
MultiCube outputs separate meshes per part, so game and animation assets come ready to edit and animate.
· “MultiCube: Compositional 3D Generation With Part-Level Semantic and Spatial Control”
Five-parameter MaxEnt beats stored QZH on 79% of HDR scenes
Maximum entropy from nine energy moments gives strictly positive irradiance and wins where light is directional.
· “Maximum Entropy Encoding of Energy-Weighted Spherical Moments”
Relational 3D DSL makes AI-shape generation stick to intent
A DSL that trades coordinates for relational layout lets an AI plan, code, and critique 3D shapes.
· “aDSL: Agentic 3D Creation via Joint Agent-Program Design”
Foveated Gaussian splatting runs 30.9x faster at human-level quality
A point-based neural renderer prunes by compute cost, relaxes the periphery, and warps one eye to reach VR frame rates.
One-by-one handle removal turns a 796k-tet mesh into 268 handles
Sparse, locally supported deformation handles keep memory low and let nonlinear elastodynamics run in real time.
· “Love Handles: Decimation for Deformation Handles with Compact Support and Low Memory Footprints”
Moving mesh vertices sharpens cloth folds with no extra vertices
A quality guard stops degenerate elements, so the same vertex budget resolves finer wrinkles and contact features.
Topology loss fixes loops and voids pixels cannot see
Adding the term cuts enclosed-void error 4.0–7.9× and loop error 2.3× at Chamfer parity.
· “Topology-Aware Differentiable Triangle-Soup Reconstruction via Persistent Homology”
Mechanism labels beat variance on hard pixels
A seven-way transport label stays stable at 64–4096 samples where variance rankings agree with themselves only 21–34%.
· “A Stable Transport-Mechanism Descriptor for Per-Pixel Rendering Difficulty”
No-training video pipeline lifts face identity by a third
Draft the motion, face-swap a few anchors, interpolate the rest — no fine-tuning needed.
· “KeyID: Decoupled Drafting and Keyframe Editing for Identity-Preserving Video Generation”
One fine-tuned video model gives any 3D character lip-synced speech
AnyTalk lifts 2D talking-head video to blendshape animation, no per-character training data required.
· “AnyTalk: Speech Animation for Arbitrary Characters Leveraging a Video Generation Model”
Implicit surfaces get exact ARAP regularization
Sampling the surface once and differentiating at sample points yields stretch and bend penalties without meshing.
· “As-Rigid-As-Possible Regularization for Implicit Surfaces”
WorldRover renders 21.9M frames with exact motion labels
Routes replay as first-person, third-person, or 360° video, pairing RGB with depth, optical flow, and point tracks.
· “WorldRover: A Scalable Synthetic Video Data Engine for World Exploration with Rich Annotations”
Sparse tangent coordinates outperform full-connectome GNNs
BrainLinear maps connectomes to a shared geometric space, keeps a few key directions, and beats 12 deep baselines at a fraction of the cost.
· “BrainLinear: A Linear Model for Brain Network Analysis in Sparse Tangent Subspaces”
Token-efficient autoregression scales 3D generation to 300 parts
Each part gets its own token budget, so assemblies stay editable and fit on one GPU where diffusion baselines fail.
Edge-by-edge odds for Jacobi sets under Gaussian uncertainty
Analytic per-edge probabilities match 10,000-run Monte Carlo within 1.8 percent, no sampling.
Grid-only training yields Lagrangian particle tracking
No particle supervision needed: the same checkpoint forecasts fields and rolls out trajectories.
· “From Fixed Grids to Moving Particles:A Transferable Latent Operator for Fluid Dynamics”
Billion-voxel volumes compressed in minutes by fixed-budget Gaussians
Gaussian count is set before training, so compression ratio and memory use are known in advance—and one model serves every visualization.
· “Fixed-Budget Gaussian Volume Encoding with Structure-Aware Allocation”
Pixels alone can narrate any gameplay video with spoken commentary
A 3x3 mosaic plus history-aware prompts turns any recording into live commentary at one ninth the image cost.
· “Content Based Video Narration of Gameplay with Vision Language Models”
One finetuning recipe turns video AI into a scene renderer
The same recipe also works in reverse, pulling scene properties back out of real footage.
· “RGBX-Next: Towards Realistic Generative Rendering from G-Buffers”
One photo becomes editable strand hair in minutes
Image-language model cues plus geometry diffusion handle ponytails, buns, and curls that learned methods miss.
· “Strand-based Hairstyle Generation via Large Reconstruction and Multimodal Models”
Subtractive 3D part generation beats additive baselines
The model repeatedly splits a part off the whole object, scoring best on part-annotated 3D geometry.
Blue-noise sampling scales to 14K in constant memory
A lattice Gibbs ensemble with a fixed-depth backward sampler makes every tile bit-identical to full-domain output.
63.9% of visualization papers miss required ethics statements
An audit of 255 visualization papers found only 6 fully reporting approval, its number, and consent.
Gaussian splats erase metal streaks in cone-beam CT
A polychromatic model with self-calibrated spectrum beats prior baselines and runs in minutes per scan.
· “Splat-based Metal Artifact Reduction in Cone-Beam CT via Polychromatic Modeling”
One RGB image predicts 3D contact and force
A model trained on simulated forces recovers hand and object pressure maps that transfer to real objects and unseen data.
· “EgoPHI: Estimating Contact and Force from Egocentric Vision”
Spline surface holds patch edges fixed during local edits
A quartic Overhauser tensor product reaches C0 and C1 seams while ruling out C2 curvature continuity.
Flow matching turns 3D anomaly detection into a density test
Projected views make every defect a low-density outlier in the Gaussian endpoint of a learned normal flow.
· “MVFM-3DAD: Multi-view Flow Matching for 3D Anomaly Detection via Density Proxy Estimation”
No controller needed: gestures drive avatars in a browser metaverse
Five-person internal trial found it usable, with a clear trade-off between precise hand navigation and immersive walking.
· “A Browser-Based Gesture-Driven Avatar Interaction Framework for Metaverse Onboarding Environments”
Fire spread simulated on real Gaussian forest scenes without meshes
Particle-driven combustion on real reconstructed forests reproduces slope, wind, and density effects.
Mocap beats every gesture AI on all four tests
Five challenged systems trail human motion, scoring near chance on conversation and meaning.
Cut-cell prior speeds skinning and improves neural baselines
Replacing voxel geodesic binding with a winding-number cut-cell graph cuts runtime and reduces deformation artifacts.
Gaussian Sculpting recovers cleaner meshes from limited views
Bi-level SDF-and-Gaussian training beats NeRF and Gaussian baselines on object-level Chamfer distance.
· “Gaussian Sculpting: End-to-End Controllable Surface Reconstruction via Field Optimization”
Layered cache turns one shaded frame into nine extrapolated frames
By filling disocclusions from real cached geometry, Amulet avoids the ghosting and blur of motion-vector warping.
· “Amulet: Frame Extrapolation Through Sparse Layered Scene Representation and Adaptive Shading”
At equal memory, hybrid BRDF beats pure neural model
Adding a 14 kB shared MLP as a correction term to GGX cuts rendering error on measured materials.
· “A Hybrid Neural-Microfacet BRDF Model for Real-Time Rendering”
Contact regions transfer real clinician bathing to a soft robot hand
A synchronized dataset of clinicians bathing people lets a soft robot hand replay the task on a mannequin.
Per-view neural branches cut 3D shape error by a quarter
A separate network per camera view, fused by self-attention, fixes the inconsistencies multi-view generators leave behind.
· “View-Adaptive Renderer for View-Consistent 2D-to-3D Generation”
Rodrigues vectors make motion stitching smooth without learned models
One quadratic solve fixes cut-and-paste jumps while keeping each segment's motion intact.
· “Smooth Motion Stitching via Laplacian Optimization in Rodrigues Vector Space”
Mapping solids: target shape becomes part of the optimization
One energy, three self-adjusting targets: fixed ellipsoid, adaptive ellipsoid, and a sea-buffered free boundary.
· “Adaptive Volumetric Parameterization of Simply Connected 3-Manifolds with Applications”
This paper presents a way to compute exact gradients through block implicit physics…
A reverse sweep of local 3x3 adjoint solves gives machine-precision gradients through the exact finite-depth Vertex Block Descent solver…
· “Differentiate the Solver, Not the Equation: Reverse-Sweep Adjoints for Block Implicit Simulation”
Compression tuning now checks whether scientific features survive
A web-based framework links rate, fidelity, and quantities-of-interest views so users pick compressors by downstream impact.
· “FZ-VIS: A Visual Analytics Framework for Quantities-of-Interest-Aware Scientific Lossy Compression”
Weak-scalability checks mis-measure data-dependent algorithms
A single-core timing probe corrects scale factors so per-core work stays constant; results agree across scaling methods.
· “Evaluating and Improving Weak Scalability Analysis of Visualization Algorithms”
Extra image sharpens single-view 3D reconstruction
Consistency-based gating picks the more informative input per view, cutting Chamfer distance by 45 percent.
· “When Does An Extra View Help? Adapting Single-View 3D Reconstruction with Extra Imagery”
A pixel-level vision model predicts how people read scatterplots
The paper argues encoding-based user studies are a dead end; image-computable models can replace them.
Exact plane clipping for Gaussian splats tops 650 FPS
A per-pixel CDF factor cuts Gaussians exactly, beating prior clipping on eight medical volumes without retraining.
· “XClipGS: Exact Half-Space Clipping for Medical Volume Gaussian Splatting”
Ray-traced sunlight pressure speeds up 15x on GPU
Vulkan backend matches reference accuracy; ignoring solar-panel rotation shifts BoxWing trajectories by 3.17 million km.
· “GPU-Accelerated Orbit Propagation with High-Fidelity Solar Radiation Pressure Modeling”
Exact trait-induced merge trees from a linear-size graph
Vertex-sampled approximations miss zero-valued features; this graph method is exact and runs in linear time.
· “Exact Computation of Trait-induced Merge Trees for Bivariate Fields”
Six photos plus 36 synthetic views beat 45 real photos for robot sims
R2S-EGO lifts held-out G1 ego-view PSNR to 19.062 dB and real sitting success to 82.5%.
· “R2S-EGO: Dual-Proxy Refinement for Sparse-Capture Real-to-Sim”
Confidence-weighted priors slash shiny-object normal error by 62%
VGGT's multi-view depth and normals fix reflective surfaces, but only when weighted by the model's confidence.
· “Confidence matters: Leveraging Multi-view Geometric Priors for GS-based Reconstruction”
ESVR represents huge sparse 3D scans as a few hundred thousand ellipsoid shapes
A method that fits differentiable ellipsoids to sparse volumetric intensity fields and renders them via per-primitive ray sampling…
Graphical abstracts, now editable: GenGA outputs SVG, not pixels
A structural-independence score measures editability, and users prefer the vector results over human-made figures.
· “GenGA: Editable and Data-Grounded Graphical Abstract Generation for Academic Papers”
Space, color, latency: why cross-reality art feels broken
24-user study of projection-AR-VR art: each mismatch hurts differently — rigid assets reveal errors first.
One static video yields a simulation-ready articulated object
Pipeline reads joint axes from boundary geometry; 42 seconds of human confirmation beats motion-based baselines.
Gaussian splats strip CT metal streaks in ~20-40 minutes
A one-scalar-per-Gaussian polychromatic model removes beam-hardening artifacts with no metal mask and keeps fine detail.
· “Splat-Based Metal Artifact Reduction in Cone-Beam CT via Compact Attenuation Modeling”
Pose errors, not sparse views, create CT streaks
Refining camera geometry during reconstruction removes streaks without blurring fine detail.
· “Revisiting Pose Sensitivity in Splat-based Computed Tomography under Sparse-view Reconstruction”
Synthetic-only model densifies sparse ToF depth across 6 datasets
One network, trained on simulated data, beats prior depth-completion methods on real sensors in 34 ms.
· “Dense Metric Depth Completion from Sparse Direct Time-of-Flight Sensors”
Multi-view Gaussians keep faces and hands sharp in avatars
The new pipeline beats generative baselines on every reported image and pose metric and runs in real time.
· “Multi-View Face and Gesture Animation with Dynamic Gaussians”
Language Gaussian splatting reaches 88.8% localization in UAV scenes
2D–3D region fusion and view-consistency filtering beat prior methods on a new four-scene UAV benchmark.
· “OutLangSplat: 3D Language Gaussian Splatting for UAV Outdoor Scenes”
Area pops out best but reads worst: channel rankings flip by task
Four-task test of seven channels shows designers need task-specific profiles, not one accuracy ranking.
· “Revisiting Channel Effectiveness: A Multi-Dimensional Evaluation with Primitive Visual Stimuli”
Seam carving shrinks images without squashing key objects
A C++ implementation of the classic content-aware resize operator, with forward energy, masks, and seam-insertion enlargement.
· “An Analysis and Implementation of Seam Carving for Content-Aware Image Resizing”
Persistent local repair lifts CAD recovery from 0.49 to 0.92
Keeping feature and step evidence, bounding edits, and reusing skills cuts geometric regression and retries.
· “TraceCAD: Trace-Guided Repair for Agentic CAD Generation”
Four seeds expose the shape of AI art uncertainty
New protocol tells compact, outlier, bimodal and diffuse readings apart—and flags when no seed matches the artwork.
