NVIDIA_WORLDGEN.md

pipelines/NVIDIA_WORLDGEN.md

NVIDIA "open-world generator" — what actually exists, mid-2026

Status: research brief, not canon. Written to resolve docs/pipeline_review/STACK_FACTS_QUESTIONS.md

Q2. Research date 2026-07-15, web-verified (training-data cutoff was Jan 2026 and NVIDIA's lineup

moved a lot since — Edify 3D alone shipped and was pulled inside that window). Every claim below is

tagged VERIFIED (checked against an NVIDIA/reputable primary source, URL inline) or INFERRED

(my synthesis bridging a gap the sources didn't state directly). Where I could not verify something

that matters, I say so instead of guessing.

0. TL;DR verdict (read this first)

**No NVIDIA product shipped as of mid-2026 is "an open-world generator" in the sense the pipeline

doc means — ingest real DEM/geodata, sculpt/populate terrain, hand back a game-ready open world.**

That thing does not exist under an NVIDIA brand. What exists instead:

1. NVIDIA's own generative-3D tooling operates at the object/prop layer, not the terrain layer.

The one piece that is real, shipped, and runs locally on a single RTX 5090 is the **AI Blueprint

for 3D Object Generation** (text prompt to Blender-ready GLB props, ~20 objects/scene) — VERIFIED,

§1.5. It populates a scene with objects. It does not create terrain, does not take a DEM, does not

output USD.

2. Nothing found does what Gaea does (heightmap sculpting, hydraulic/thermal erosion, terrain

authoring) under NVIDIA's own generative label. UE 5.8's new native Mesh Terrain system

(Epic's own feature, not NVIDIA's) is the closest 2026 development in that specific slot —

VERIFIED, §2.

3. Cesium remains the real-world-geodata leg, unchanged. NVIDIA doesn't ingest DEM/satellite data

itself; Cesium for Omniverse is the connector that streams Cesium's own tiles (terrain,

imagery, OSM buildings — sourced by Cesium, not NVIDIA) into an Omniverse/USD stage, as an

*optional alternative staging point* to feeding Cesium into UE directly. It is a Cesium product

that happens to plug into NVIDIA's platform, not an NVIDIA geodata product. VERIFIED, §1.8.

4. **NVIDIA's actual mid-2026 push (Cosmos 3, Earth-2, the GTC 2026 Omniverse announcements) is almost

entirely industrial physical-AI / robotics / autonomous-vehicle / climate infrastructure, not game

content creation.** Cosmos generates video/action data for robot and self-driving-car training, not

3D meshes for games. Earth-2 generates climate/weather simulation, not terrain meshes. VERIFIED,

§1.1, §1.6.

5. **The things that actually match the "describe it, get a populated open world" shape a solo dev

might have half-remembered are third-party, not NVIDIA** — Tencent's OpenUSD-based world pipeline,

Promethean AI (text-to-populated-3D-scene, acquired by a studio in 2025), and Overworld

(VC-funded startup, real-time local diffusion world model, demoed AT NVIDIA's GTC but is not an

NVIDIA product). These are real and worth knowing about, but "the NVIDIA open-world generator" is

the wrong name for any of them. VERIFIED, §1.9.

Practical recommendation for the pipeline correction pass: retire the premise that a single

NVIDIA product replaces any leg of the Cesium / Gaea2Unreal / SRTM chain. Keep that chain exactly as

designed. The one NVIDIA piece worth actually wiring in is the 3D Object Generation Blueprint, as a

local, RTX-5090, prop/object-population step that runs after terrain exists — i.e., it plugs in

next to UE-PCG population, not in place of Cesium/Gaea/SRTM. Full input contract in §4.

---

1. The lineup, product by product

1.1 Cosmos — VERIFIED, not a game-asset tool

NVIDIA Cosmos 3 ("omnimodel") natively understands and generates text, images, video, ambient

sound, and actions with physics-grounded outputs

(NVIDIA Cosmos product page;

Cosmos 3 launch, NVIDIA Newsroom, June 2026;

HPCwire coverage).

The stated purpose is training data and simulation for robots and autonomous vehicles

(warehouses, factories, roads) — every primary source frames this as physical-AI/robotics, never

game development or terrain generation. Model matrix: Cosmos-Predict2.5 ships as 2B and 14B variants

(model matrix;

GitHub); the 2B Text2Image variant is confirmed

runnable on one RTX 5090 (32GB), the 14B variant needs ~64GB+ VRAM (effectively dual 80GB GPUs or a

single H200) — INFERRED/aggregated from secondary GPU-requirement writeups, not an NVIDIA primary

source, flagged accordingly. Licensed under OpenMDW 1.1 (Linux Foundation) per the Cosmos product

page. **Bottom line: even the locally-runnable Cosmos variant outputs video/images, not meshes or

USD — it cannot be a terrain or asset generator for a game pipeline no matter how you wire it.**

1.2 Omniverse — VERIFIED, repositioned away from creative-tool framing

Omniverse is not a generator; it's a USD-based platform/SDK (Kit) used to build apps like USD

Composer and USD Explorer

(Kit overview).

As of ~May 2026 NVIDIA dropped the subscription requirement — Omniverse is now free for development,

production, and redistribution, with no AI Enterprise subscription required for that free tier

(StorageReview coverage).

GTC 2026's Omniverse announcements (DSX Blueprint, Mega Omniverse Blueprint, OpenClaw) are

industrial-digital-twin / robot-fleet / AI-factory framing, not game content

(Into the Omniverse, GTC 2026).

Game-adjacent Omniverse usage that does exist and is real: Omniverse Audio2Face for AI facial

animation, adopted by GSC for S.T.A.L.K.E.R. 2 and Fort Solis; Tencent Games adopted Omniverse to

unify USD-based asset pipelines

(Omniverse game-dev blog; note this

source reads like an older/evergreen NVIDIA blog post reused across years — treat the Tencent/GSC

claims as VERIFIED-as-stated but not necessarily current-2026 news). None of this is world/terrain

generation — it's animation tooling and asset-pipeline unification.

1.3 Edify 3D — VERIFIED discontinued as a direct NVIDIA offering, with a caveat

Edify 3D (announced GTC 2024, trained on Shutterstock's licensed 3D/image library, powering

Shutterstock's and Getty's own generative-3D services) — search evidence indicates the direct NVIDIA

NIM preview for Edify was discontinued around June 2025. I could not independently re-verify this

against a live NVIDIA source in this pass (the canonical research.nvidia.com/labs/dir/edify-3d/

page now 301-redirects to research.nvidia.com/labs/cosmos-lab/edify-3d/, which reads as Edify having

been folded under the Cosmos-lab umbrella rather than cleanly killed — INFERRED, worth a direct

follow-up check before treating as settled). What is clear: NVIDIA's own 2025-2026 developer-facing

3D-object tooling (the Blueprint in §1.5) is built on Microsoft's TRELLIS, not Edify — i.e.

NVIDIA's internal direction moved away from Edify as the generative-3D backbone. Whether

Shutterstock's/Getty's own Edify-branded commercial services are still live is unverified in this

research pass; if the pipeline ever depends on Edify specifically, re-check

shutterstock.com directly rather than trusting this brief.

1.4 NIM microservices for OpenUSD — VERIFIED as announced, current status partly unconfirmed

The July 2024 announcement bundled several generative NIMs for OpenUSD workflows: USD Search

(text/image query over existing 3D libraries — GA/preview on the NVIDIA API catalog), USD Layout

(assemble a USD scene from text prompts via spatial intelligence), USD SmartMaterial (auto-apply

materials), fVDB Mesh Generation (point-cloud to USD mesh, explicitly cloud-rendered via

Omniverse Cloud APIs), fVDB Physics Super-Res

(NVIDIA Newsroom announcement;

developer blog).

USD Layout in particular is the one that sounds closest to "world generator" (spatial-intelligence

scene assembly from text) — but it was "coming soon" as of the 2024 announcement and **I could not

confirm a current mid-2026 shipped/GA status for USD Layout, USD SmartMaterial, or fVDB Mesh

Generation in this research pass.** Do not assume any of the three are live and usable — check

build.nvidia.com/explore/visual-design directly before writing a Translation Doc against them.

USD Search is the one confirmed-available piece of this set, and it is a search tool over

existing assets, not a scene/terrain generator.

1.5 The AI Blueprint for 3D Object Generation — VERIFIED, the one real local match

This is the most concrete, currently-shipping, RTX-5090-local NVIDIA generative-3D offering found in

this research pass.

(NVIDIA blog announcement;

GitHub repo;

build.nvidia.com listing).

Full input contract detail is in §4. In one line: text prompt in, up to 20 individual GLB-format 3D

objects out, auto-exported to Blender — an object/prop-population tool, explicitly not a terrain

or whole-world generator (the source article's own framing: "lets 3D artists create up to 20 3D

objects to prototype a scene").

1.6 Earth-2 — VERIFIED wrong domain

Earth-2 is NVIDIA's climate/weather digital twin: kilometer-scale atmospheric simulation (storms,

cloud cover, global weather), using the CorrDiff generative model, visualized through Omniverse

Cloud's Graphics Delivery Network

(Earth-2 product page;

kilometer-scale climate model blog).

Every source describes outputs as weather/climate simulation datasets and visualizations — none

describes a terrain mesh, heightmap, or DEM export product. Its early adopters (The Weather Company,

Taiwan Weather Administration) are forecasting organizations, not game studios. **This is not a

terrain generator for games under any reading of the sources.**

1.7 Rendering tech that is not generation

DLSS 5's "3D-guided neural rendering" (announced GTC 2026, targeted Fall 2026) reconstructs pixels

in real time using structured 3D data as an AI conditioning signal — Jensen Huang's framing is

"fusing structured data in virtual worlds with generative AI's probabilistic computing"

(TechRadar GTC 2026 live coverage).

This is an upscaling/rendering technique applied to a scene that already exists, not a content

generator. Likewise RTX Kit (ReSTIR PT), Mega Geometry's foliage system (path-traced dense

vegetation, partnered with CD PROJEKT RED for The Witcher), and RTX Remix (remastering old

fixed-function game assets) are all rendering/geometry-processing technologies, not world/terrain

generators

(NVIDIA RTX game-dev blog).

1.8 Cesium for Omniverse — VERIFIED, the real geodata connector (and it isn't NVIDIA's)

Cesium for Omniverse streams Cesium's 3D Tiles — Cesium World Terrain, Bing imagery, Cesium OSM

Buildings, high-res photogrammetry — directly into an Omniverse/USD stage, with LOD streaming and

caching

(Cesium product page;

NVIDIA developer blog on the integration;

GitHub). It's free/open-source, Apache 2.0. This

directly answers the pipeline doc's "does it ingest geodata natively" question about the NVIDIA

piece: **no NVIDIA generative model ingests DEM/geodata directly — Cesium does that job exactly as

today, and Cesium-for-Omniverse is only relevant if the pipeline decides to stage terrain through

Omniverse/USD before UE import, which is an optional detour, not a required or NVIDIA-native one.**

A solo dev targeting UE5.8 directly can just as well use Cesium for Unreal (the existing

UE-native Cesium plugin) and skip Omniverse entirely for this leg — INFERRED as the simpler path

given no unique NVIDIA capability is gained by routing through Omniverse for geodata specifically.

1.9 Adjacent products that are NOT NVIDIA (the likely source of the ambiguity)

These plausibly explain where "the NVIDIA open-world generator" phrase came from — they were shown

at NVIDIA's own events, or run on NVIDIA hardware, but they are third-party products:

• Overworld — an independent, VC-funded ($4.5M pre-seed, Kindred Ventures et al.) studio building

real-time, local-first diffusion world models for "living, adaptive worlds." Demoed at both GDC

2026 and NVIDIA's GTC 2026, runs on consumer hardware

(GlobeNewswire;

Waypoint-1.5 release).

No confirmed NVIDIA partnership beyond presenting at the conference — it is not built or sold by

NVIDIA.

• Promethean AI — text-to-populated-3D-environment tool (describe a scene, get placed/scaled/

textured assets with spatial logic), reported acquired by a studio in 2025. Independent company,

not NVIDIA.

• Tencent's OpenUSD-based production pipeline — Tencent Games built a proprietary

Unreal-Engine-plus-OpenUSD asset pipeline (described at a GTC 2024 session,

"Build Games in the Cloud: How Tencent Streamlines 3D Content Production With OpenUSD"),

using Omniverse/USD as infrastructure. This is Tencent's own studio tooling built on top of

NVIDIA's platform, not an NVIDIA product a solo dev can download. (I found a secondary-source

reference to a Tencent "procedural content generation framework" demoed at GDC 2026 with striking

numbers — 10 km² scene generation in under 10 seconds — but could not independently confirm the

product name or details against a primary Tencent/NVIDIA source; treat that specific claim as

unverified and do not cite the "Genesis Scene" name without further confirmation.)

None of these three should be called "the NVIDIA open-world generator" — if Josh has one of these

specifically in mind (most likely candidate given the RTX-5090-local framing: Overworld, since it

explicitly targets "everyday hardware" and was the one with a live local demo), that's a

genuinely-different research task (evaluating a startup's product, not an NVIDIA SDK) and should

be flagged back rather than folded into this brief's verdict.

---

2. USD / Omniverse -> Unreal Engine interchange — the practical mechanics (VERIFIED)

This matters regardless of which generative piece (if any) gets adopted, because the pipeline's

target engine is pinned to UE 5.8 (per docs/PIPELINE_REVIEW_MISSION.md / Josh, 2026-07-15).

• **The NVIDIA Omniverse Connector for Unreal Engine supports UE 4.27 through 5.3 only — it does NOT

support 5.4 or newer**

(Omniverse Connect requirements page).

For UE 5.4+, NVIDIA's own documentation tells you to use Epic's native USD plugin instead of

the Omniverse Connector. **This means for a UE 5.8 target, the Omniverse Connector plugin is

simply not an option — the interchange path has to run through Epic's own USD support, with or

without any NVIDIA generative piece in the loop.**

• Epic's native OpenUSD support in UE 5.8 ships as the USDImporter plugin and the

InterchangeOpenUSD translator

(USD Importer, UE 5.8 docs;

Universal Scene Description (USD) in Unreal Engine, UE 5.8 docs).

Epic's own docs flag USD support as Beta — workable, but caution before shipping on it.

I could not pull the deeper workflow docs (round-trip/live-sync specifics) past the landing page in

this pass; if the correction pass needs exact live-sync behavior, read the full UE 5.8 USD guide

pages directly rather than trusting a secondhand summary.

• UE 5.8 also ships an experimental native Mesh Terrain system — full 3D sculpting without

heightfield limitations, a genuinely new (Epic-built, not NVIDIA) terrain-authoring option that

didn't exist in earlier UE5 versions

(Unreal Engine 5.8 feature breakdown).

Worth the pipeline doc's own attention independent of anything NVIDIA — it's a second, native option

alongside Gaea2Unreal for the terrain-sculpting leg.

• **The one NVIDIA generative-3D tool that's actually usable today (§1.5, the 3D Object Generation

Blueprint) sidesteps the whole USD-compatibility question** — it outputs GLB (glTF binary), and

UE5.8 has a mature native glTF/Interchange importer independent of the USD plugin's Beta status.

This is a practical point in its favor for a solo dev who wants the shortest, most stable interchange

path.

---

3. Verdict — ranked, what each replaces, what it can't do

Ranked by "how plausible is this as what a solo dev with a 5090 would actually reach for":

1. AI Blueprint for 3D Object Generation (TRELLIS NIM) — REAL, SHIPPED, LOCAL-5090-CONFIRMED.

Replaces: nothing in the Cesium/Gaea2Unreal/SRTM terrain chain. Slots in as an **object/prop

population step**, functionally adjacent to (not a replacement for) UE-PCG population — i.e. it's

another way to generate the props/set-dressing that populate a region once terrain exists, usable

alongside or instead of hand-placed/PCG-scattered assets for prototyping. Cannot: sculpt terrain,

take a DEM/heightmap as input, output USD, or generate anything at world/region scale (it's capped

at ~20 objects per prompt, single-object-at-a-time generation under the hood).

2. Omniverse + Cesium for Omniverse — REAL, SHIPPED, FREE, but an *optional detour*, not a

generator. Replaces: nothing that Cesium-for-Unreal doesn't already do directly for a UE-only

pipeline. The only reason to route through it is if the pipeline wants a USD-based intermediate

staging/collaboration layer (e.g. for cross-DCC asset assembly with Maya/3ds Max/Houdini before UE

import) — not needed for a solo dev going straight Cesium-to-UE. Cannot: generate new terrain or

content; it only stages/streams what Cesium already has.

3. Everything else in §1 (Cosmos, Earth-2, USD Layout/SmartMaterial/fVDB, Edify, DLSS 5/RTX Kit) —

REAL but wrong-domain or wrong-maturity for this pipeline. Cosmos and Earth-2 are physical-AI/

climate infrastructure that happen to share NVIDIA's brand and some Omniverse plumbing with game

tooling, not usable substitutes for terrain or asset generation. The USD Layout/SmartMaterial/fVDB

NIMs are the closest on paper to "describe a scene, get a populated 3D world," but their current

shipped status is unconfirmed — do not architect around them without a direct build.nvidia.com

check first. Edify is likely discontinued/absorbed. DLSS 5/RTX Kit are rendering tech, not content

generation, full stop.

What this means for the terrain_pipeline_type enum proposed in STACK_FACTS_QUESTIONS.md Q2: the

draft recommendation floated {nvidia_worldgen, cesium_feed, srtm_feed, hybrid} on the assumption

that an nvidia_worldgen value would replace the Gaea leg. Based on this research, there is no

verified NVIDIA product to put behind that enum value for terrain sculpting/generation — the honest

enum is closer to {cesium_feed, srtm_feed, gaea_sculpt, ue_mesh_terrain, nvidia_object_populate},

where nvidia_object_populate is the TRELLIS-NIM Blueprint doing prop/object population on top of

whichever terrain path (Gaea or UE 5.8's new native Mesh Terrain) is chosen — not a peer alternative

to Cesium/SRTM/Gaea, but a downstream step after them.

---

4. Input contract detail — AI Blueprint for 3D Object Generation (for the Translation Doc)

This is the one piece worth writing a Translation Doc against. Everything here is VERIFIED from the

NVIDIA blog post

and the GitHub repo directly.

Hardware prerequisites

• GPU: RTX 5090, RTX 5080, RTX 4090, RTX 4080, or RTX 6000 Ada (consumer/prosumer RTX only — no

datacenter GPU required)

• System RAM: 48 GB minimum
• OS: Windows 10/11 only
• CUDA Toolkit 12.8
• ~50 GB free disk space for model weights
• A free HuggingFace account + access token (models are pulled from HuggingFace at setup)
• Python via Miniconda, Git + Git LFS, Visual Studio Build Tools (native dependency build chain)
• Blender 4.2+ (optional — only needed for the Blender add-on integration path)

Pipeline stages (the actual generation chain under the hood)

1. User provides one text prompt describing a scene idea (e.g. "night market," "sunny day at the

beach") via a chat interface.

2. An LLM (Qwen3-4B or Llama 3.1 8B, run locally via NIM) brainstorms up to ~20 candidate objects

that would belong in that scene, each with its own generation sub-prompt.

3. NVIDIA SANA (a fast text-to-image model) generates a 2D preview image for each candidate

object from its sub-prompt.

4. Microsoft TRELLIS (via its NIM microservice, image-to-3D) converts each 2D preview into an

actual 3D object.

5. Each object exports as an individual GLB (glTF binary) file.

6. The companion Blender add-on ("3D Object Generation" + "Asset Importer") auto-imports the GLB set

into a live Blender scene for review/refinement.

Invocation (as documented)

• Install/enable the Blender add-ons, point the add-on at the Blueprint's local install path, open

the N-panel in the 3D Viewport, click "Start Services," wait for READY status, then open the

"3D Object Generation UI" to enter the scene prompt.

• The Blueprint can also run headless/standalone from its own local chat UI without Blender, per the

repo's service-based architecture (NIM containers exposed locally).

Output contract

• Format: GLB per object (no USD output from this specific tool).
• Cardinality: up to ~20 discrete objects per single scene prompt — object-level, not a merged scene

file, not terrain, not a level layout.

• Texturing: baked into the GLB (TRELLIS generates textured meshes, not separate material graphs to

hand-author).

Licensing (commercial game dev)

• Blueprint orchestration code: Apache 2.0 (NVIDIA's repo).
• TRELLIS model + code: MIT — explicitly commercial-use-permitting per Microsoft's own repo

(microsoft/TRELLIS). Caveat: two rendering dependencies

used internally (nvdiffrast, nvdiffrec) carry their own separate licenses — worth a one-time check

before shipping a product that bundles them, though for a dev-time-only asset-generation tool

(never redistributed inside the shipped game) this is unlikely to matter.

• NIM microservices generally (Qwen/Llama/SANA/TRELLIS containers): free under the NVIDIA Developer

Program for "research, application development, and experimentation on up to 16 GPUs"; an NVIDIA

AI Enterprise license (~$4,500/GPU/year, or ~$1/GPU-hour cloud) is required for production use,

defined by NVIDIA as anything beyond dev/test/research/evaluation, specifically "activity serving

real end-users"

(NIM product FAQ). **Reasoned application to this

pipeline (INFERRED, not NVIDIA-stated in so many words): generating art assets locally during

development — where the NIM containers never serve a live player and the output is just baked GLB

files committed into the game's asset folder — reads as development/research use under the free

tier. If the pipeline ever calls a NIM container at runtime inside the shipped game (e.g. live

per-player procedural generation), that flips to production use and needs the Enterprise license.**

This distinction is significant enough that it's worth a direct confirmation from NVIDIA's actual

EULA text (not secondary sources) before the correction pass treats it as settled — flag as

UNVERIFIED-AT-EULA-TEXT-LEVEL.

---

5. Direct answer to STACK_FACTS_QUESTIONS.md Q2

Which NVIDIA product is "the open-world generator" (Earth-2 class? the GTC world-gen models? NIM

microservices?), and does it replace all three legs or only sculpting/generation?

None of the above, cleanly. Earth-2 is climate simulation (wrong domain). The GTC world-gen

models (Cosmos) are physical-AI/robotics video simulation (wrong domain, wrong output type — video,

not geometry). The NIM microservices split into: USD Search (confirmed, but it's a search tool, not

a generator), USD Layout/SmartMaterial/fVDB (announced, current status unconfirmed — don't build on

these without checking build.nvidia.com first), and the one that's actually real and running locally

today, the 3D Object Generation Blueprint (TRELLIS NIM) — which operates at the object/prop

layer, not terrain, and replaces none of Cesium/Gaea2Unreal/SRTM. It's a fourth, additive step

(object population) that runs after terrain exists, not a substitute for any of the three terrain

legs. **Recommend correcting the pipeline docs to keep the Cesium/Gaea2Unreal/SRTM chain exactly as

designed**, and separately wire in the TRELLIS-NIM Blueprint as a local RTX-5090 prop-population tool

alongside UE-PCG — with the nvidia_worldgen enum value replaced by something like

nvidia_object_populate scoped honestly to what it actually does.

If Josh had a specific different product in mind when he wrote "the NVIDIA open-world generator" —

most plausibly Overworld (§1.9), given it's the one real-time, local-hardware, "generate a

playable world" demo that actually showed up at NVIDIA's own GTC 2026 — that's worth a direct

follow-up question, since Overworld is not an NVIDIA product and evaluating it is a different research

task (startup product with its own SDK/licensing/maturity profile, not an NVIDIA-blessed pipeline

component).

---

Sources (all fetched/searched 2026-07-15)

• NVIDIA Cosmos product page
• Cosmos 3 launch — NVIDIA Newsroom
• Cosmos 3 — HPCwire/AIwire coverage
• Cosmos-Predict2.5 model matrix
• nvidia-cosmos GitHub org
• cosmos-predict2.5 GitHub repo
• Omniverse Connect — Unreal Engine requirements
• Omniverse Kit overview
• Omniverse made free for production — StorageReview
• Into the Omniverse — GTC 2026 virtual worlds / physical AI
• Omniverse accelerates game content creation (Audio2Face, Tencent)
• Edify Shutterstock/Getty announcement
• Edify 3D generative AI / fine-tuning blog
• Shutterstock Generative 3D press release
• research.nvidia.com Edify-3D page (redirect notice)
• NVIDIA OpenUSD generative NIMs announcement (July 2024)
• Integrate generative AI into OpenUSD workflows — developer blog
• NVIDIA Expands Omniverse With Generative Physical AI (SimReady)
• AI Blueprint for 3D Object Generation — NVIDIA blog
• 3D Object Generation Blueprint — GitHub
• 3D Object Generation Blueprint — build.nvidia.com
• 3D-Guided Generative AI Blueprint (Blender + FLUX NIM)
• Microsoft TRELLIS — GitHub (MIT license)
• Microsoft TRELLIS.2 — GitHub
• NVIDIA Earth-2 product page
• Earth-2 kilometer-scale climate model blog
• NVIDIA GTC 2026 keynote live coverage — TechRadar
• NVIDIA RTX innovations for game development — developer blog
• GeForce GDC 2026 announcements
• Cesium for Omniverse — product page
• Leverage 3D geospatial data with Cesium — NVIDIA developer blog
• cesium-omniverse — GitHub
• USD Importer — UE 5.8 docs
• Universal Scene Description (USD) in Unreal Engine — UE 5.8 docs
• Unreal Engine 5.8 feature breakdown (Mesh Terrain)
• NIM product FAQ / licensing
• Access to NVIDIA NIM free to Developer Program members
• Overworld GDC/GTC 2026 announcement
• Overworld Waypoint-1.5 release
• Tencent GTC 2024 session — OpenUSD 3D content production

Generated by harness/site/structure_site.py — the URL path is the repo path. review root