3D Spatial Intelligence Open-Source Toolbox — 955 methods, 18 skills, MCP Server, SplatVerse Studio
The plugin will be installed here. Keep web if you are unsure.
npx -y @deepseek-ai/dsh plugin --profile web add github:jaccen/Awesome-Gaussian-Skills#8b7f60a5f32fd407c523343d7d5768958fa7b33d
Compatibility and provenance
Awesome Gaussian Skills is published as awesome-gaussian-skills and currently resolves to version 0.9.11. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.
Other awesome lists give you paper titles. We give you paper titles + an AI toolkit that makes you faster. The canonical lists catalog the ecosystem for humans; this repo makes it consumable by AI agents — let your AI read the papers for you.
Then ask your Agent: "Compare 3DGS and 2DGS rendering formulations"
Integrations
This repo ships as both AI agent skills and an MCP server, so it works across agent runtimes and MCP clients — the knowledge base is consumable by your AI agent, not just by you:
Client
How to use
Claude Code / OpenClaw
/plugin marketplace add jaccen/Awesome-Gaussian-Skills → install gaussian-research, gaussian-engineering, or gaussian-domain (18 skills in 3 plugins); or load skills/ directly (see CLAUDE.md); MCP server via the config below
DSH (DeepSeek Harness)
dsh plugin --profile web add jaccen/Awesome-Gaussian-Skills — the 18 skills + MCP renderer auto-mount via cordis.patch.yml
A full open-source technical book, built around 3D Gaussian Splatting as the spine and weaving together spatial intelligence and embodied intelligence into one closed loop: representation → perception → planning → action.
What's inside — 13 chapters + hands-on prelude & capture SOP, every method name anchored to this repo's real data (955 Methods, 26 categories, 18 skills), zero fabrication
#
Chapter
Focus
引言
Why this book
Why 3DGS is the key puzzle piece of Physical AI
前
Landing prelude: should you, can you
Tech-selection matrix, eight cases where you should NOT use 3DGS, three reading paths
CH 01
NeRF → 3DGS: A paradigm leap
Explicit vs implicit, the three innovations, the alpha-compositing formula
CH 02
The math & engineering core
Anisotropic Gaussians, differentiable rasterization, adaptive density control, CUDA
操
Data-capture SOP
Device selection, trajectory design, seven exposure hard rules, scale calibration —the first step that decides your quality ceiling
Each chapter ends with hands-on exercises and links back to the repo's method tables, references/, and skills —so reading the book and doing the engineering are one seamless flow.
Highlights you won't find in a paper list:
The six schools of world models (2026 taxonomy) and where 3DGS sits as the only representation that is simultaneously renderable, differentiable, and editable.
How GS-World, ManiGaussian, and OrbiSim turn 3DGS into a differentiable simulation engine.
A three-tier model of robot spatial memory (geometric → appearance → semantic) and where current GS-SLAM actually stands.
Why we wrote it (and how it relates to this repo)
This book is the narrative layer over the repo's data layer. The repo gives you 955 method names, abstracts, and 18 engineering skills —but not the through-line that connects them. The book supplies that through-line: it argues why 3DGS became the bridge between spatial intelligence and embodied intelligence, and walks every chapter back to concrete methods and skills you can use today. Read the book to understand the map; use the repo to ship the territory.
What's New (Oct 2026)
Latest Update (Oct 15, 2026): v0.9.11 — Research Expert Team Deep Expansion: Methodology Core + 15 Master-Paper Studies ⭐
Pure skill-content release — no carrier changes (955 methods / 26 categories / 18 skills stay).
research-expert-team grew from a nine-role router into a two-layer methodology asset:
Methodology core (program layer) — executable procedures distilled from 15 top-cited
exemplars: topic-selection engine (M1-M4 + five opening-sentence variants), argumentation
skeletons (T1-T3 + representation-paper techniques B.5), 16 experiment patterns (C1-C16),
eight reviewer must-ask questions with preempts (D1-D8), nine value amplifiers.
Previous Update (Oct 13, 2026): v0.9.10 — AIGC Content Creation Category + Research Expert Team Skill ⭐
New category AIGC Content Creation (22 methods; taxonomy 25 → 26): the AI-generated-content
creation layer over 3DGS — text/image → 3DGS scenes and assets, generative scene reconstruction,
stylization, artistic fonts, avatars. Anchored by DiffSplat, Dragen3D, PSGS, Lyra, CSD,
SplatFont3D plus 2026 entries (VirSqueezer, CGGS, UniWorld-View, OmniStyle-INR,
Flex4DHuman, LightHarmony3D); the pre-existing CGGS entry was completed with its arXiv ID
(2607.03819) and re-categorized. New skill research-expert-team (17 → 18 skills): a nine-role
full-cycle research team — idea evaluation, innovation mining, simulated peer review, academic
polishing, AIGC-trace hygiene (publication hygiene), framework & venue compliance, strict venue-format
writing, experiment design, experiment analysis — with method-selection tables for the three frontier
directions (Ontology / Super Agents / AIGC Content Creation). Library 934 → 955; all data carriers
re-verified at 955, name sets identical. See changelog.
Previous Update (Oct 13, 2026): v0.9.9 — Knowledge-Base Expansion: 900 → 934, Two New Skills ⭐
34 new arXiv-verified methods admitted, taking the library from 900 to 934, with the
taxonomy growing 23 → 25 categories: two new cross-cutting categories — Ontology & Knowledge
Representation (12 methods) and Super Agents & Agentic AI (20 methods). Two new skills
shipped — ontology (3D scene graphs, spatial ontologies, neuro-symbolic reasoning) and
super-agent (multi-agent LLM system design, 3D/spatial agent integration) — skill count
15 → 17. CGS-SLAM name-collision remediated (colliding newcomer dropped; original
IROS 2025 entry restored). All data carriers re-verified at 934, name sets compared entry
by entry; site-wide caliber unified to 934 / 25 / 17. See changelog.
Targeted reinforcement of the thinnest category. 4 new arXiv-verified 3DGS-based World
Model / Spatial Intelligence methods admitted, taking the library from 896 to 900. The
World Models & Spatial Intelligence category grew from 8 → 12 methods:
HY-World 2.0, F3DGS (federated multi-agent world modeling), 4DGS-WAM (object-centric
world action model), Cloth-Splatting (3DGS cloth state estimation). Two missing arXiv IDs
backfilled (Holi-Spatial → 2603.07660, Spatial-TTT → 2603.12255); two entries flagged
unverified pending source check (APEIRIA, S²AM3D). GSMem found already in library
under Embodied AI & Robotics — not duplicated. See changelog.
Manual admission of 15 new arXiv-verified 3DGS methods, taking the library from
881 to 896 (+15, +1.7%). Admissions span 9 categories: Acceleration (H2-Ray, TileSkipper,
SplatQ), Feed-Forward (CompactSplat, DeltaSplat, DensiTok), Embodied AI & Robotics
(Splat2Act, 2DGS-Planner, NovaSplat), plus OuroWorld (Dynamic & 4D), PAM-ToD (HDR & Relighting),
FlyMark (Security), GSCV (Compression & Streaming), OpenSplatGraph (Language & Semantic),
PCAsplat (Surface & Rendering). Every admitted method name is self-stated in its own paper's
title/abstract (R9 protocol) and arXiv-verified. 12 candidate papers were rejected for lacking a
self-stated method name, being benchmark-only, NeRF (not 3DGS), or using 3DGS merely as a tool. All data
carriers (methods.json / CSV / methods.html / abstracts.js / reference doc) re-verified at 896,
name sets compared entry by entry. No method name was fabricated.
Derived from systematic gap analysis across the 955-method knowledge base.
Target venues: TVCG / CGF / CAD / T-RO / IJCV / ACM TOG.
I-01. Part-Aware Alpha-Compositing for Articulated Objects
Problem: Standard alpha-compositing causes color bleeding at part boundaries of articulated objects. ULF-Loc (CVPR 2026) exposed this feature bias, but no rendering-formulation-level fix exists.
Approach: Extend alpha-compositing with part-aware opacity modulation: C(θ) = Σ Tᵢ · αᵢ · ω_{p(i)}(θ) · cᵢ(θ), where ω penalizes penetration and joint violations, making inter-part penetration regions automatically transparent.
Path: 1) Build on gsplat rasterizer. 2) Add FK layer for articulated objects (URDF). 3) Compute penetration/joint violation via SDF. 4) Train on Articulate-100.
Target: SIGGRAPH / ACM TOG / TVCG
I-02. Geometry-Consistent Flow World Model for Manipulation
Problem: Flow-based world models (RoboFlow4D) predict dense 3D flows but lack geometric consistency —predicted flows can violate object rigidity and physical constraints.
Approach: Couple 3D flow prediction with scene graph constraints: rigidity loss for static objects, articulation loss for joints, support-relation constraints for stacking.
Path: 1) Extend RoboFlow4D. 2) Scene graph parser via OpenMask3D. 3) Geometric regularizer. 4) Train on LIBERO + RoboCasa.
Target: IJCV / T-RO / RSS
I-03. Multi-Scale Occupancy-Gaussian Bidirectional Bridge for Driving
Problem: Occupancy (SparseWorld, DOV) is the driving world model standard; 3DGS provides superior rendering. No differentiable bridge exists between them.
Approach: Bidirectional conversion: Occ→ DGS (learned position+scale predictor from occupancy+semantics) and 3DGS→ Occ (differentiable sparse convolution pooling). Joint backbone for unified prediction+rendering.
Each skill follows the SKILL.md standard, compatible with Claude Code (.claude/), Cursor (.cursor/rules/), Windsurf, and other AI Agent frameworks.
SplatVerse Studio: Short Video Creation
SplatVerse Studio integrates 3D Gaussian Splatting with a short-drama pipeline powered by the Toonflow engine — from text scripts to 3DGS-rendered video scenes (7-step pipeline: Script Adaptation → Storyboard → Toonflow Sync → TTS → Video Gen → FFmpeg Compose → Publish).
📖 Full documentation & integration guide (incl. MoneyPrinterTurbo):studio/README.md
@misc{awesome-gaussian-skills,
author = {jaccen},
title = {Awesome Gaussian Skills: 3D Spatial Intelligence Open-Source Toolbox for 3D Gaussian Splatting Research},
year = {2026},
url = {https://github.com/jaccen/Awesome-Gaussian-Skills},
note = {955 arXiv-verified methods across 26 categories, 18 AI agent skills, and an MCP server with 34 tools}
}
Ecosystem & Acknowledgments
Complementary projects — we recommend all of them:
Routine daily arXiv scan admitted 13 new arXiv-verified 3DGS methods, taking the library from
859 to 872 (+13, +1.5%). Admissions span Surface & Rendering, Feed-Forward, Large-Scale,
Compression, Sparse-View, Cross-Domain, Embodied AI & Robotics, Dynamic, and Language & Semantic
categories (incl. LiteTex-GS, GrapeSplat, VDGS, CRP-GS, NormLift). 7 candidate papers
were rejected for failing the verification protocol — benchmarks without a self-stated method name,
descriptive-only paper titles, or 3DGS used merely as a tool. All data carriers
(methods.json / CSV / methods.html / abstracts.js / reference doc) re-verified at 872,
name sets compared entry by entry. No method name was fabricated; the library's
"every entry arXiv-verifiable" guarantee holds.
Knowledge base expanded from 700 to 859 verified methods (+159, +22.7%) through a multi-channel arXiv
harvest over the last 150 days. Six query families (cs.CV / cs.GR / cs.RO / full-corpus /
abs:"3D Gaussian Splatting" / ti:"Gaussian Splatting") returned 395 candidates; after verification
162 were admitted. Also in this release:
Verification protocol (R1–R6): every admission requires the paper's self-stated method name,
reproduced in its own abstract; multi-word names must contain a proper-noun token — this alone rejected
67 descriptions masquerading as method names (Capacity-Controlled, Scene-Level). 114 borderline
candidates were read individually; 25 adjudicated in, the rest dropped.
Accuracy repairs: 57 reference links had a displayed arXiv ID differing from the hyperlinked one
(all had the correct ID displayed, wrong link). 13 mis-attributed entries deleted — e.g.
Dynamic 3D Gaussians pointed at 2309.13114, which is actually a magnetic spin-glass physics paper,
and GS-LRM-v2 pointed at DOF-GS. All now verified against the arXiv API.
AIGC traces removed from 29 text files; 21 historically corrupted JPEGs restored and re-stripped safely.
README_CN.md recovered from double-encoding corruption that had mangled every Chinese string.
Full carrier sync — methods.json / CSV / methods.html / abstracts.js / reference doc
all verified at 881, name sets compared entry by entry rather than by count alone.
Adjudication wave (same day). A second pass resolved every name carrying a derived suffix or
colliding with another entry: Scaffold-GS+ and GaussianSplatting-SLAM-v2 deleted (no such paper
exists), Luminance-GS++ renamed to its self-declared base name, and Spark 2.0's description
corrected — it is World Labs' open-source WebGL2 splat renderer, not an NVIDIA robotics simulator.
Two independent papers both self-name GeoGS-SLAM (2607.07452 and 2607.11184, disjoint author lists);
they are two methods, not a sequel, and are now disambiguated. 862 → 859.
Latest update (Sep 20): v0.9.2 —Scene Agent + Mapmost Engineering Wave: 5 New Methods + 6 Skill Enhancements. Knowledge base expanded from 692→ 97 methods. Added 5 scene-generation Agent papers: SceneSmith (ICML 2026 Spotlight, MIT CSAIL —hierarchical designer-critic-orchestrator VLM agents for simulation-ready indoor scenes), Scenethesis (ICLR 2026, NVIDIA —training-free LLM+vision agentic framework), SceneGenAgent (ACL 2025 —LLM coding agent for industrial scenes), SceneConductor (arXiv —multi-agent orchestration from single image), SceneAssistant (arXiv —visual-feedback agent for open-vocabulary scene generation). Injected Mapmost production engineering knowledge into 4 skill files: 3DGS individualization (单体化) in best-practices.md, LOD streaming + seam-fix + adaptive partition training in best-practices.md, 3DGS+3DTiles multi-source fusion in gis-toolchain.md, LUT color grading in tech-stack.md, Mapmost as Digital Twin player with 100km² benchmark in industry-landscape.md. Categories updated: Generation +4, Embodied AI & Robotics +1.
Previous (Sep 20): v0.9.1 —ECCV 2026 Wave: 14 New Methods. Knowledge base expanded from 678→ 92 methods. All 14 new entries arXiv-verified or ECCV-official: Wat3R (underwater 3D geometry without annotations), GeoNVS (geometry-grounded video diffusion with GS-Adapter), Edit3r (instant 3D editing from unposed images), SA-ResGS (first residual learning for 3DGS in next-best-view selection), GaussianLens (on-demand localized high-res reconstruction), Deformable Triangle Splatting (non-convex primitive with per-edge control points), Neural Harmonic Textures (per-primitive harmonic feature decoding), CubicSplat (error-bounded differentiable vector graphics), TetraSDF (analytic isosurface extraction), Heat Kernel Textures (ECCV 2026 Best Paper —geodesic Gaussians), GRF-Recon (long-sequence feed-forward with global ray-field), PriSplat (distractor-free 3DGS via multi-view propagation), Per-View Distractor Filtering (training-free feed-forward artifact removal), GS-Underwater (controlled cross-regime underwater benchmark study). Categories updated: Surface & Rendering +5, Feed-Forward +4, Cross-Domain +2, Robustness +2, Editing +1, Embodied AI & Robotics +1. All data carriers in sync.
Previous (Sep 16): v0.9.0 —Data Accuracy Release. Full-library audit of every entry against the arXiv API: each arXiv ID re-verified reachable and matched to its paper (method name must appear in the title/abstract). 181 entries failed verification and were removed, 56 wrong-ID entries re-linked to their real papers, 5 renamed to the papers' self-declared method names, plus 1 duplicate removed. The count is now honest instead of inflated: 678 methods, 23 categories, 568 arXiv-verified (83%), 130 code links (19%). Per-entry audit evidence: docs/data-accuracy-audit-2026-09-16.md. See changelog/2026-09-16-v5.md.
Previous (Sep 4): v0.8.3 —Dynamic Scene Survey Integration: 23 New Methods. Knowledge base expanded from 801→ 19 verified methods. Added 23 arXiv-verified dynamic scene reconstruction methods from survey paper (连振晗 et al., J CAD & CG, Jan 2026), spanning 3 categories: deformation field (LoopGaussian, CoGS, CD-GS, PGED, GPS-Gaussian, MoDGS, MoDec-GS, SpectroMotion, BARD-GS, GauFRE, ReconDreamer++), sparse-control deformation (SP-GS, Video-3DGS, SplineGS), 4D Gaussian primitive (Real-time 4DGS, PVG, 4D-rotor GS), and per-frame inter-frame transfer (3DGStream, Dual-GS, STC-GS, IGS, GFlow, DynOMo). Dynamic & 4D category: 75→ 7. New references/dynamic-datasets.md (11 datasets) and expanded benchmark-data.md (MS-SSIM/VMAF/FID metrics + 4 performance comparison tables). Experiment planner skill updated with dynamic scene design guide. See changelog/2026-09-04.md.
Previous (Sep 2): v0.8.2 —Daily Update: 11 New Methods. Knowledge base expanded from 790→ 01 verified methods. New additions span 7 categories: BRF-GS (hyperspectral BRDF modeling with 3DGS), SMG (ECCV 2026, semantic motion graph for dynamic GS), VCAR (ACM MM 2026, training-free 3DGS segmentation), RealCAD (real-world image to CAD reconstruction), ObjectSplat (object-level mesh splatting), CapFrame (ECCV 2026, text-instructed viewpoint localization), Amortized Anchor Refinement (continuous-time 4D reconstruction), ATGS (SIGGRAPH 2026, anchored temporal GS for long volumetric video), CausalSplat (hierarchical causal reasoning in 3DGS), VoroTracing (623 FPS differentiable ray tracing), Lucida (Real-to-Sim composable scene modeling). All arXiv IDs verified, data CI passed. See changelog/2026-09-02.md.
Previous (Aug 23): v0.8.1 —Daily Update: 7 New Methods. Knowledge base expanded from 783→ 90 verified methods. New additions span 6 categories: LEGO (ECCV 2026, hierarchical language GS with LLM spatial reasoning), OutLangSplat (UAV outdoor open-vocabulary 3D language GS), ESVR (IEEE VIS 2026, 3D ellipsoid sparse volume rendering with 4 orders of magnitude compression), TRACE-GS (sparse-view 3DGS via privileged geometric conditioning), RORA (single-video-to-articulated-object pipeline with Unreal Engine deployment), OVOW (ECCV 2026, monocular video to instance-level 4D mesh for physics simulation), Super-Gaussian (interactive 3DGS scene editing with VR NLI visualization). All arXiv IDs verified, data CI passed. See changelog/2026-08-23.md.
Previous (Aug 7): v0.8.0 —Platform Upgrade (P0+P1+P2). Knowledge layer: single source of truth (data/methods.json, 783 methods, 23 categories) with data CI; 5 fabricated entries purged; 14 arXiv-verified frontier methods added. Capability layer: true-3DGS render loop (gsplat via HTTP-served PLY), server-authoritative scene persistence, real PLY/SPLAT export, 5 distinct prune strategies, grid-accelerated ray query, runtime arg validation, WS origin allowlist, 21 unit tests. Platform layer: Benchmark arena (bench/), skill orchestration contracts (skills/_contracts/), Router manifest loader (scripts/router_load.py). 13 core MCP tools (all real) + 13 experimental (gated by INCLUDE_EXPERIMENTAL=1). See changelog/2026-08-07.md.
