WORLD_MODELS_LANDSCAPE.md

pipelines/WORLD_MODELS_LANDSCAPE.md

World models & generative worlds — the Q3 2026 landscape

Status: research brief, not canon. Research date 2026-07-15, web-verified (training-data cutoff was

January 2026 and this whole category moved fast inside that window — Odyssey, Decart, Runway, and

Alibaba all shipped or announced new products since). Every claim is tagged VERIFIED (checked

against a primary source — the vendor's own site/docs/blog/GitHub/HuggingFace — or corroborated by

≥2 independent secondary outlets, URL inline) or INFERRED (my synthesis bridging a gap the

sources didn't state directly) or UNVERIFIED (a claim I found but could not independently

confirm in this pass — flagged rather than trusted). Where sources conflict, both readings are shown.

This is the third of three parallel research lanes spun up over docs/pipeline_review/STACK_FACTS_QUESTIONS.md

Q2 (the "NVIDIA open-world generator" re-examination): a dedicated Cosmos-3 deep dive and a

UE-export-bridges lane are the other two. **This doc gives NVIDIA Cosmos 3 landscape-level treatment

only** (§2.A, §5) — where it appears, read it as corroborating context, not the final word; a

COSMOS3_DEEP_DIVE.md-class doc, if/when it lands, is the authority on Cosmos-3-specific

game-dev-applicability. Likewise §3 (export mechanics) goes as deep as a landscape survey needs but

defers to a dedicated bridges-lane doc for exact interchange mechanics if one lands. Sibling reading:

tech_research/NVIDIA_WORLDGEN.md §1.9 first name-dropped Overworld/Promethean/Tencent's OpenUSD

pipeline in passing — this doc is the requested deeper, wider follow-up. RESEARCHED_STACK.md

Chain 2 is the already-ratified world-chain (Copernicus DEM → GDAL → Gaea/UE Mesh Terrain → PCG

Biome Core populate); nothing here overrides it — §7 says explicitly what does and doesn't touch it.

---

0. TL;DR verdict

**The field splits into two paradigms that do fundamentally different jobs, and almost everything

loud in the press (Genie 3, Odyssey, Overworld, Decart, Cosmos 3) is the wrong paradigm for an

authored, 79-node, hand-canon RPG.** Those systems generate a video stream frame-by-frame,

conditioned on your inputs, with no persistent exportable geometry underneath — you cannot pull a

mesh, a heightmap, or a level out of them. They are consumer experiences and robotics/synthetic-data

engines, not content pipelines. A smaller, less-hyped cluster — **World Labs' Marble, Tencent's

HY-World 2.0, Meta's WorldGen (research-only), Alibaba's Happy Oyster** — instead outputs *persistent,

editable 3D scene data* (mesh, Gaussian splats, point clouds) that a game engine can actually import.

That second cluster is the one worth evaluating as pipeline tooling; the first is worth knowing about

but not architecting around.

**For scene assembly specifically — the nearer-term, more relevant question given our region-page

schema — the field is thinner than the world-model hype suggests.** Promethean AI is real, still

shipping, still UE-integrated, but (a) its public plugin repo is pinned to UE 5.3 with no confirmed

5.8 support, and (b) like every system in this space it takes a free-text prompt, not a structured

schema — which cuts directly against this project's own ruled principle that Claude Code drives

NeoStack schema-bound, never freehand (STACK_FACTS_QUESTIONS.md Q1). **UE's own native PCG

(Production-Ready since 5.7, GPU-Compute path, PCG Biome Core, DataTable ingestion) plus the

already-owned NeoStack plugin already cover this job better than any third-party "describe a scene"

tool** — nothing found in this pass beats that combination for a schema-bound pipeline. Epic's own

UE6 (Early Access targeted late 2027) is building exactly this capability natively, with Claude named

as a first-class MCP integration — the one system in the "upcoming" bucket worth actually designing

around, on a multi-year horizon.

Nothing in this landscape ingests a region page directly. Every world model and every

scene-assembly tool takes prose/images/panoramas as input, not typed rows (zone catalogs, site rows

with terrain/functions, per-site vril data). That gap is not a missing product to wait for — it's the

adapter layer this project already plans to write by hand (Claude Code reading the region page and

driving PCG/NeoStack), and no vendor roadmap found in this pass closes it for us.

---

1. The paradigm question — two families, not one race

Every piece of press coverage treats "world models" as a single race with a leaderboard. It isn't.

There are two fundamentally different output contracts, and which one a product has determines

whether it can touch our pipeline at all.

Family A — interactive world-model games (the world IS the model). The system is a video

diffusion or autoregressive model that predicts the next frame given the current state and your

input (keyboard/mouse/text/action tokens). There is no scene graph, no mesh, no persistent geometry

anywhere — "the environment" only exists as the model's running latent state, and it decays: every

system in this family loses coherence after a few minutes (Genie 3: "a few minutes... no more"

— DeepMind Genie 3 blog).

You cannot export a level from this family. You can only *watch and steer* it, live, inside its own

player. Genie 3/Project Genie, Odyssey's current interactive-video line, Overworld, Decart's Oasis/

Lucy, Microsoft's Muse/WHAM, Runway's GWM-1, and (per this pass's evidence — see §2.A) NVIDIA Cosmos 3

all belong here.

Family B — authored-engine tooling (world models that hand you a scene, not a stream). The

system still uses generative diffusion/reconstruction internally, but its *output* is a scene

description in a standard interchange format — a textured mesh, a Gaussian-splat point cloud, a

depth/normal map set — that a game engine imports once and then owns forever, editable in the DCC

tools a studio already uses. World Labs' Marble is explicit about this distinction in its own

marketing: "the output is not a rendered video but a digital 3D scene file that can be modified,

populated with objects, and experienced from any angle"

(TechCrunch, World Labs x Autodesk).

Tencent frames HY-World 2.0 the same way, explicitly contrasting itself against the first family:

"every world model before HY-World 2.0, including Genie 3... NVIDIA Cosmos, and Tencent's own

HY-World 1.5, generates pixel-level video. In contrast, HY-World 2.0... produce\[s\] actual 3D

geometry, not video" (HY-World 2.0 GitHub;

corroborated by BuildFastWithAI's landscape comparison).

This is the only family that can be a *pipeline node* rather than a standalone experience.

Why the split is a hard line, not a spectrum, for this project specifically. Humanity is

77 chapters of hand-authored canon over real-Earth regions, with a typed event graph (beats → forks →

state), a 20-section region-page template, and a ruled principle that generation is always

schema-bound, never freehand (STACK_FACTS_QUESTIONS.md Q1: "every action... is MATCHED TO OUR

SCHEMA... no freehand generation"). Family A cannot serve this at any maturity level — even a perfect

Family-A model would hand you an ungrounded, non-reproducible, non-editable video that has to be

manually rebuilt into an authored scene anyway, so its "generation" work is thrown away. Family B can

serve it, in principle, exactly as much as any other asset-generation tool: as raw material a

human/Claude Code re-authors against the schema, never as a drop-in final level. Unity's own CEO made

the industry-level version of this argument in response to Project Genie: **"We believe world models

are going to be a source of inspiration and assets for creators, but that they're not in any way going

to replace game engines. They are complementary, not duplicative,"** and "Unity is not an interactive

video generator; it's a real-time 3D execution platform designed to build once and then run everywhere"

(Matthew Bromberg, Unity CEO, Q4 2026 earnings call — The Motley Fool,

corroborated by Nasdaq

and The Globe and Mail).

Read Unreal for Unity in that quote and it's exactly our situation.

---

2. The full roster

2.0 Comparison table

2.A Family A in more depth — the interactive-world-model cluster

Google DeepMind Genie 3 / Project Genie. The most press-visible entrant. Text prompt →

real-time explorable environment at 24fps/720p, holding coherence "for a few minutes, no more," with

navigation/camera control and "promptable events" (weather, appearances) but "no fine-grained physical

manipulation yet" (DeepMind Genie 3; Google blog, Project Genie).

Access is gated entirely behind the $200/month Google AI Ultra tier, 18+, US-first with a

globally-widening but feature-limited rollout (Engadget;

the-decoder on the 2026 I/O pricing restructure).

There is no public API and no export path of any kind — it is a consumer toy, not a pipeline

component, at any price point.

Odyssey. London AI lab, now the highest-valued pure-play in this space ($1.45B, June 2026,

Amazon/AWS-backed — TechCrunch).

The 2026 news cycle is dominated by its interactive-video line — Starchild-1 (multimodal,

audio-synced), Agora-1 (multiplayer, demoed as a "GoldenEye-style" real-time multi-agent world),

Odyssey-2 Pro (720p/22fps) — which is explicitly Family A: frame-by-frame video, "no game engine in

sight." Free API signup (email only) plus a web sandbox at odyssey.ml. Notably, Odyssey's earlier

"Explorer" product (per odyssey.ml/introducing-explorer)

claimed real mesh export — USD, FBX, glTF — putting it in Family B; whether Explorer is still a live,

current product alongside the video line in mid-2026, or was superseded, was not confirmed in this

pass and would need a direct site check before relying on it either way.

Overworld (formerly Wayfarer Labs). The one genuinely local-first, open-weight entry in Family A.

Waypoint-1.5, released April 2026: 1.2-3B-class models, 720p/60fps, running on "everyday hardware"

rather than a GPU cluster (GlobeNewswire).

The HuggingFace org page states MIT license directly on the model card — worth flagging that

press materials elsewhere say "Apache 2.0"; both are commercially-permissive so the practical

implication is the same, but the two don't literally agree and the primary HF source should govern.

Still Family A (video/pixel output, no mesh, no export) — genuinely interesting as the most

accessible/cheapest system to *experiment* with, not as something with any current path into a UE

scene.

Decart. Two products: Oasis 3, a photorealistic driving-simulation world model built with

Etched, explicitly for autonomous-vehicle/robotics training (geometry-aware, multi-camera, <200ms

latency, $0.02/sec via API — TechCrunch),

and Lucy 2.0, a real-time video-transformation model (editing, not generation from scratch).

Neither has any relationship to game-world authoring; $300M raised May 2026 at a ~$4B valuation with

NVIDIA, Adobe Ventures, and Toyota Ventures among the backers

(SiliconANGLE).

Microsoft Muse / WHAM / WHAMM. Built with Xbox Game Studios' Ninja Theory, published in *Nature*,

explicitly framed as a gameplay-ideation tool, not a production pipeline component — Microsoft's

own materials stress it augments rather than replaces developers

(Microsoft Research).

WHAMM (the real-time extension) hits 10+fps versus WHAM-1.6B's ~1fps; a live Quake II demo runs in

Copilot Labs. Weights and the demonstrator are open-sourced on Azure AI Foundry and HuggingFace. No

export path — it's a prediction/ideation model over an existing Xbox title's action space, not a

general 3D generator.

Runway GWM-1. Three variants — GWM Worlds (explorable environments), GWM Avatars (conversational

characters), GWM Robotics (synthetic training data) — built on Gen-4.5, autoregressive, real-time,

24fps/720p (Runway Research). Research

preview as of December 2025 with broader access "expected early 2026"; this pass did not confirm

whether that access has actually opened by mid-2026. Family A throughout — no mesh/export path found.

NVIDIA Cosmos 3. Landscape-level facts only (full treatment belongs to the dedicated deep-dive):

launched ~May 31/June 2026 as "the first omni-model with native reasoning, world and action

generation" (NVIDIA developer blog;

MarkTechPost).

Two shipped sizes: Cosmos3-Nano (16B) and Cosmos3-Super (64B). The "world and action

generation" claim is real and VERIFIED — but the actual output class is images/video/synchronized

audio/action-conditioned rollouts, the same Family-A shape as Genie 3, not persistent editable

geometry (HuggingFace launch blog).

On hardware: Cosmos3-Nano at 16B parameters is ~32GB in BF16 — that consumes essentially the

entirety of an RTX 5090's 32GB with no headroom for activations/context, and NVIDIA's own target

hardware framing for Nano is the workstation-class RTX PRO 6000 (96GB), not a consumer card

(Spheron Blog +

WaveSpeed) — INFERRED that unquantized

Nano is not comfortably 5090-local; a quantized (FP8/INT8) variant could plausibly change this but

none was confirmed in this pass. Every primary NVIDIA source frames Cosmos 3 for robots/autonomous

vehicles/physical-AI, never game content. **Net read consistent with NVIDIA_WORLDGEN.md's original

verdict**: Cosmos 3 is real, is a genuine "world model," and does generate "worlds" in the video sense

— but it is Family A (video/action rollout, wrong output type for a UE content pipeline) and

wrong-domain-positioned (robotics, not games) by NVIDIA's own framing. Treat the dedicated

Cosmos-3 deep-dive as authoritative if it finds a different bottom line.

ByteDance and Meta's non-game work, briefly: ByteDance has not shipped a world model as of

this research pass — the Seed team's stated goal is Genie-3 parity by end of 2026, with a dedicated

sub-team (led by Fan Haoqi) explicitly aimed at "entertainment and games" as a target application

(36Kr; ChinaBizInsider).

Worth a re-check at end-2026. Meta's V-JEPA 2 (robotics prediction) and Code World Model (32B,

code-generation reasoning) both use "world model" in the technical/predictive sense and are **not

game-content-relevant at all** — the actual game-relevant Meta project is WorldGen (§2.B).

2.B Family B in more depth — the exportable-3D cluster

World Labs Marble is the most mature product in this family. Founded by Fei-Fei Li on a

"spatial intelligence" thesis, it turns text/image/video/panorama/coarse-layout input into an

explorable, physically-stable 3D world, and — critically — actually exports it: Gaussian splats

(with a collision mesh) for high-fidelity visuals, or standard mesh (FBX 2018+/glTF 2.0) for a

game-engine-native path

(techcrunch.com coverage).

A $1B round closed February 18 2026, anchored by $200M specifically from Autodesk (plus AMD,

Emerson Collective, Fidelity, NVIDIA, Sea) explicitly to bring world models into professional 3D

workflows — Maya/3ds Max-class DCC integration is the stated direction but **not yet shipped or

detailed** as of this pass

(TechCrunch).

Pricing is a hard gate for us: **Free (4 generations) and Standard ($20/mo, 12 generations) carry NO

commercial rights at all** — Pro ($35/mo, 25 generations, adds commercial rights + scene expansion) is

the minimum tier that could legally touch a shipped game; Max is $95/mo/75 generations

(Marble pricing). VR-compatible (Vision Pro, Quest 3). Also has

a growing robotics-simulation use case via NVIDIA Isaac Sim integration (PLY/GLB → USD/USDZ path) —

notable because it confirms Marble's output is genuinely engine-agnostic structured 3D data, not a

locked-in format.

Tencent HY-World 2.0 is the technically closest thing to a true open, self-hostable,

game-engine-ready world generator found in this entire pass. Open-sourced April 16, 2026, full weights

and code, with an explicit multi-stage pipeline (Panorama Generation → Trajectory Planning "WorldNav"

→ World Expansion "WorldStereo 2.0" → World Composition "WorldMirror 2.0" + Splatting Learning) that

turns text or a single image into mesh/3DGS/point-cloud output the vendor states imports directly into

Blender, Unity, Unreal Engine, and Isaac Sim

(GitHub; 3D-model tech report PDF).

It currently ranks #1 on Stanford's WorldScore benchmark among open-source world models and is

described as benchmark-comparable to Marble. The license is the open question: the HuggingFace

model card names it tencent-hy-world-2.0-community — a bespoke Tencent license, not plain Apache

2.0 or MIT, despite loose secondary coverage calling it "open license." Every sibling Tencent Hunyuan

release checked in this pass (HunyuanVideo, Hunyuan3D-2, HunyuanWorld-Mirror) carries an EU/UK/South

Korea exclusion plus a MAU-based commercial gate (100M MAU for HunyuanVideo-class products, as low as

1M MAU for HunyuanWorld-Mirror) under a "community"-named license

(HunyuanVideo LICENSE; DeepWiki on HunyuanWorld-Mirror commercial licensing) —

this is directly the same restriction pattern already flagged for Hunyuan3D-2.1 in

LOCAL_3D_ASSET_GEN.md / STACK_FACTS_QUESTIONS.md Q3, which Josh already ruled around (territory-

restricted tools stay non-load-bearing in the shipped-asset path). Worth noting Tencent's own naming

generation ("Hy3") reportedly dropped territorial restrictions in its July 6, 2026 full release

under genuine Apache 2.0 — but that data point is about the broader Hunyuan-3-generation family, not

confirmed to cover HY-World-2.0's specific -community license, so treat HY-World-2.0's terms as

unresolved pending a direct LICENSE-file read rather than assuming either the old restrictive

pattern or the new open one.

Alibaba Happy Oyster (1.0), shipped June 2026: 3D environments plus interactive/rewindable video,

explicitly marketed for film/game/visual-experience production

(Bloomberg).

Genuinely new and under-covered in English-language sources found in this pass — worth a dedicated

follow-up if it becomes a real candidate, but not deep enough here to rank with confidence.

Meta WorldGen is the actual game-relevant Meta project (distinct from V-JEPA2/CWM, §2.A): a

research system that turns a single text prompt into an interactive, navigable 3D world the paper

claims is compatible with Unity and Unreal "without the need for additional conversions"

(Meta Quest Blog;

arXiv:2511.16825). Not available to developers — research

only. Worth tracking for a 6-12-month re-check, not usable today.

CSM AI (Common Sense Machines), Cube 2.0: MIT-founded (Cambridge, MA) team building "controllable

3D game-world and video generation from any input" (text/image/sketch) via a proprietary diffusion +

NeRF-style pipeline (csm.ai). This name surfaced late in the research pass and was

not run through the same depth as the others — access model, pricing, and export format are

genuinely unconfirmed here; flag as a name worth a dedicated look, not yet evaluated.

2.C Scene-assembly / layout AI — a different category from world models

This is the category the task specifically flagged as the nearer-term win, and it deserves its own

bucket because it's conceptually distinct from both families above: these tools don't generate a

*world*, they *arrange existing assets* inside a scene per a description — closer to an AI level

designer than an AI world model.

Promethean AI. Still an active, shipping product in 2026 (Indie $29/mo, Professional $89.99/mo,

Enterprise custom — prometheanai.com). Describe a scene in plain

English ("a small sci-fi lab with a workbench and scattered tools") and it proposes layout, prop

placement, and materials pulled from your own asset library, learning from your prior choices over

time. Open-source plugins exist for Unreal, Unity, 3ds Max, Maya, and Blender, explicitly

adaptable/compilable if your exact version isn't pre-tested. The concrete risk for us: the public

GitHub repo is literally titled "PrometheanAI Unreal 5.3 plugin"

(GitHub), and no source found in this

pass confirms UE 5.8 compatibility — for our hard UE 5.8 target this needs direct verification (or

plugin porting work) before it could be relied on, the same caution class already applied to the

NVIDIA Omniverse Connector's UE-5.3 cap in NVIDIA_WORLDGEN.md §2.

UE-native PCG. Not a third-party product at all — Epic's own Procedural Content Generation

framework, Production-Ready since 5.7 (confirmed again in this pass, matching RESEARCHED_STACK.md

Chain 2 step 5). At GDC 2026, Epic demonstrated a 4km×4km procedural jungle (the "Electric Dreams"

project) with zero lines of code; 5.7/5.8 added a GPU-Compute path (2x perf improvement on point

distribution/density calculations) and PCG Biome Core for ecosystem-zone management across large open

worlds (blog.imseankim.com GDC 2026 writeup;

Creative Bloq).

Separately, community tooling (the "Unreal MCP Server") lets PCG graphs be driven by natural-language

prompts via Model Context Protocol — conceptually the same shape as the Claude-Code-drives-NeoStack

principle already ruled for this project, just via a different plugin

(StraySpark).

Unity's own direction (comparison context only — we're on UE, not Unity). Unity Muse is

deprecated, replaced by "Unity AI," an in-editor Gemini-powered assistant that launched open beta

May 4, 2026. Unity's positioning is explicitly "controllable generation" rather than full

world-generation — reinforcing §1's paradigm argument from the competing engine's own mouth

(iconpolls.com Unity AI review).

NVIDIA's own scene-assembly angle is already fully covered in NVIDIA_WORLDGEN.md §1.4-1.5 —

USD Layout (text-prompt scene assembly via "spatial intelligence") was announced in 2024 but its

current 2026 shipped status is still unconfirmed, while the AI Blueprint for 3D Object Generation

(TRELLIS NIM) is real, shipped, and RTX-5090-local, but generates individual props, not scene layouts.

Nothing new surfaced in this pass changes that verdict — not re-litigated here.

---

3. Export paths to Unreal — the practical mechanics

Three real paths exist across everything surveyed; a fourth (video) is a dead end for pipeline

purposes.

1. Mesh/glTF/FBX — the cleanest path. World Labs Marble (FBX 2018+/glTF 2.0), Tencent HY-World 2.0

(mesh output, vendor-stated direct Unity/Unreal import), Meta WorldGen (vendor-claimed

no-conversion-needed Unity/Unreal compatibility, research-only), and Odyssey's Explorer line

(USD/FBX/glTF, current status unconfirmed) all land here. UE 5.8 has a mature, native glTF/

Interchange importer — this is the same practical advantage already identified for the NVIDIA

TRELLIS Blueprint in NVIDIA_WORLDGEN.md §2: no Beta-flagged plugin, no version-cap risk, shortest

path to something usable.

2. Gaussian splats (PLY/SPZ) — the higher-fidelity but more fragile path, used by Marble's

splat export and any Luma-class real-world capture. **UE 5.8 ships with no native/first-party

Gaussian Splatting module** — confirmed directly against the UE 5.8 release notes,

which don't mention it as a built-in feature. The path runs entirely through third-party plugins:

the Luma AI plugin (free, distributed via Fab/Marketplace, drag-and-drop .ply/.luma import,

the most production-relevant first-party-adjacent option), XVerse (free, GitHub, less actively

maintained), NanoGS (free, "Nanite-style" splat rendering,

CG Channel),

plus several smaller community plugins (GSplatForUnrealEngine, MLSLabsGaussianSplattingRenderer-UE,

unreal-splat) and paid options (Postshot, Volinga, a paid "3D Gaussian Plugin" on the Marketplace).

PLY is the de facto standard interchange format across this ecosystem. A mature mid-2026 plugin

ecosystem reportedly hits 60fps for scenes under ~1M Gaussians on mid-range GPUs — workable, but

every plugin in this space is third-party and none is Epic-blessed as of 5.8.

3. USD — already fully covered in NVIDIA_WORLDGEN.md §2: Epic's native USD support in UE 5.8

ships (USDImporter plugin + InterchangeOpenUSD translator) but is flagged Beta by Epic's own

docs; the NVIDIA Omniverse Connector caps at UE 5.3 and is not an option for a 5.8 target. World

Labs offers a PLY/GLB→USD/USDZ conversion path aimed at Isaac Sim (robotics), not UE, for whatever

it's worth as a data point on ecosystem direction. Not re-derived further here.

4. Video (no export) — Genie 3, Odyssey's current interactive line, Overworld, Decart, Runway

GWM-1, Microsoft Muse/WHAM, and NVIDIA Cosmos 3 all produce this. There is no mechanical bridge

from a video stream to an authored UE scene; the only "export" is a human watching and manually

rebuilding what they saw, which defeats the purpose of using the tool as a pipeline step at all.

---

4. Scene-assembly AI grounded against our actual region-page schema

The task asked specifically whether any of this can consume our region pages — the zone catalog and

site rows our pipeline actually produces. Grounding this against the real template

(_source/01_Tier_1_Foundation/T1_Region_Page_Template [ACTIVE v1.0].md, currently the Phase-C

revision base per docs/pipeline_review/GAP_MAP.md finding #10):

• The "zone catalog" is Section 17 ("Environment areas required per environment: name, purpose,

square meters, key visual elements, vril site present, build priority" + "Interior spaces... name,

purpose, dimensions, ceiling height, lighting character, audio character, key interactive

elements"), cross-fed by Section 1's terrain/build data (USGS/SRTM dataset, Gaea import

settings, UE5 landscape scale, water plane height, time-period terrain modifications).

• The "site rows with terrain/functions" are Section 1's vril-site sub-entries (name, type,

recharge rate, cultural significance, access, UE5 build notes) plus Section 8's territory/faction

fields and Section 9's mineral-deposit rows — all typed, all structured, none of it free prose.

No system surveyed in §2 ingests this shape of input. Every world model and every scene-assembly

tool (Marble, HY-World 2.0, Promethean AI, Unity AI, UE's own MCP-driven PCG control) takes a text

prompt, an image, or a panorama — none takes a typed row set with named fields and cross-references.

This is not a product gap likely to close soon; it's a fundamentally different design point (these

are consumer/creative-professional tools built for "describe what you want," not enterprise-schema

ingestion). Two ways this could still be used, both requiring a human/Claude-Code adapter step in the

middle rather than a direct pipe:

1. Prose-bridge for previz/reference only: render a region's Section 1/17 rows into a natural-

language scene description (Claude Code already does exactly this kind of transformation elsewhere

in the pipeline) and feed *that* to Marble or HY-World 2.0 to get a fast explorable reference/

blockout to eyeball before hand-building the real UE scene. This treats the AI output as

inspiration, never as final content — consistent with the do-not-invent-canon discipline and the

schema-bound generation principle; the human-authored region page stays the source of truth, and

the AI-generated reference gets thrown away or heavily reworked, not imported wholesale.

2. Direct schema-to-PCG, no third-party tool needed: UE's own PCG framework already has

first-class DataTable ingestion (RESEARCHED_STACK.md Chain 3: "Load Data Table → attribute sets

is the direct mechanism for registry rows becoming populated zones"). Since our region-page rows

already land in registries as CSV/DataTable rows, **the schema-to-scene path already exists and

needs no new vendor product** — it needs the region-page template v2 revision (Phase C, already

planned) and PCG graphs authored against it, driven by NeoStack + Claude Code per the already-ruled

architecture principle. This is the load-bearing finding of this section: **the "scene-assembly AI"

this project needs is not a product to shop for — it is NeoStack + UE-native PCG, already owned and

already the ruled plan**, and every third-party "describe a scene" tool surveyed here is either

redundant with that (Promethean AI) or usable only as disposable previz (Marble, HY-World 2.0).

---

5. Licensing / access quick-reference

---

6. Upcoming — worth designing around vs. worth ignoring (2026H2/2027)

Worth designing around, on its real timeline: Unreal Engine 6. Announced at State of Unreal

2026 (Unreal Fest Chicago), targeting Early Access at the end of 2027 with full commercial release

12-18 months after that (realistically late 2028/mid-2029) — matching RESEARCHED_STACK.md's existing

note that "UE 5.8 CONFIRMED current and terminal for UE5... UE6 EA ~late 2027." Its headline AI feature

is exactly our §4 scene-assembly problem, natively: a text prompt furnishes a room, then expands it

into a full city with roads/buildings/lighting/VFX auto-placed and fully manually editable afterward,

built on a new Verse-based persistent-world architecture, with **MCP protocol integration naming

Claude, Gemini, and Codex as first-class model integrations**

(Engadget;

GIGAZINE). This is the single

most relevant "authored-engine tooling" system on any horizon for this project specifically — but it's

2+ years out at Early Access alone, so it's a "track the beta, don't wait for it" item, not a current

plan input.

Worth a 6-12-month re-check:

• ByteDance's world model — explicit Genie-3-parity target by end of 2026, with a games/

entertainment-focused sub-team; re-check access/output-type once shipped.

• World Labs × Autodesk DCC integration — $200M committed, "entertainment use cases" named as the

starting focus, but no shipped Maya/3ds Max bridge yet as of this pass; worth checking again once

something ships, since it would directly extend §4's previz-bridge option.

• Meta WorldGen — the one Family-B research system with an explicit "no conversion needed"

engine-compatibility claim; watch for a developer-access announcement.

• Odyssey Explorer's current status — resolve whether the mesh-export product line is still live

alongside the video-focused news cycle.

• Any first-party Gaussian Splatting module for UE — Epic's own roadmap signals interest but

nothing shipped through 5.8; if it lands, it removes the third-party-plugin fragility flagged in §3.

• HY-World 2.0's actual license text — a five-minute direct read of the repo's LICENSE file that

this pass could not complete; worth doing before any benchmark-day evaluation that includes it.

Worth actively ignoring for this project's purposes: everything else in Family A (§2.A) at any

future version bump. A faster, longer-coherence, higher-resolution Genie 4 or Cosmos 4 would still be

the wrong output type — better video is not closer to an exportable UE scene, it's just better video.

Don't let a splashy future demo reopen this question; the paradigm mismatch is structural, not a

maturity gap that improves with scale.

---

7. What serves our pipeline — ranked

TODAY — already the right call, this pass just reinforces it:

1. UE-native PCG (Production-Ready since 5.7) + PCG Biome Core, reading region-page rows via

DataTable ingestion, per the already-ratified RESEARCHED_STACK.md Chain 2 step 5. Nothing surveyed

here beats it for schema-bound scene assembly, and it needs no new subscription, license check, or

version-compatibility risk.

2. NeoStack (owned, lifetime) driven by Claude Code, schema-bound — per the already-ruled

architecture principle (STACK_FACTS_QUESTIONS.md Q1) — is this project's actual answer to "AI

scene assembly." §4's finding is that this ruling was already correct; no third-party tool surveyed

in this pass changes that.

3. World Labs Marble and/or Tencent HY-World 2.0, scoped narrowly to disposable previz/reference

(never final content): feed a Claude-Code-authored prose bridge of a region's Section 1/17 rows in,

get a fast explorable blockout to eyeball before hand-authoring the real scene. If adopted, Marble

needs the Pro tier ($35/mo) the moment any output touches a shippable asset even indirectly (to

avoid any commercial-rights ambiguity); HY-World 2.0 needs its license read directly first per §5.

Neither should be treated as more than a reference aid given neither ingests our schema.

6-12 MONTHS — re-check, don't build against yet: ByteDance's world model once shipped; World Labs

× Autodesk's DCC bridge once it exists in a concrete form; Meta WorldGen if it opens to developers;

Promethean AI's actual UE 5.8 compatibility (a single direct check would resolve this cheaply, but

per §7.1 it's likely redundant with NeoStack even if compatible); a first-party UE Gaussian Splatting

module if Epic ships one.

PARADIGM-MISMATCH — track for awareness, do not architect around: Genie 3/Project Genie, Odyssey's

interactive-video line, Overworld, Decart, Runway GWM-1, Microsoft Muse/WHAM, NVIDIA Cosmos 3 (pending

the dedicated deep-dive's final word), Niantic's Large Geospatial Model. All are real, several are

genuinely impressive, and all are the wrong output shape (video/action rollout, or real-world capture

with no indie access path) for an authored, schema-bound, 79-node canon RPG. UE6's native AI world-

population (§6) is the one system in this entire landscape that could eventually close the gap these

can't — and it's Epic's own tooling, arriving on Epic's own multi-year timeline, not a vendor product

to adopt now.

---

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

• Overworld — GlobeNewswire, Waypoint research preview
• Overworld — Waypoint-1.5 release, GlobeNewswire
• Overworld — HuggingFace org page (primary, license)
• Overworld — blog index
• Google DeepMind — Genie 3
• Google DeepMind — Genie 3 blog, "a new frontier for world models"
• Google — Project Genie announcement
• Engadget — Project Genie hands-on
• the-decoder — Google AI subscription restructure, I/O 2026
• Odyssey — official site
• Odyssey — Explorer announcement
• Odyssey — Series B announcement
• TechCrunch — Odyssey $1.45B valuation, Amazon-backed
• TechCrunch — Odyssey video-to-interactive-world model
• Decart AI — official site
• Decart AI — Oasis
• TechCrunch — Decart Oasis 3 photorealistic driving
• SiliconANGLE — Decart $300M raise
• World Labs — Marble pricing
• World Labs — Marble Unreal export docs
• World Labs — official site, Labs
• World Labs — robotics case study (Isaac Sim)
• World Labs — funding announcement blog
• TechCrunch — World Labs x Autodesk $200M / $1B round
• TechCrunch — World Labs Marble commercial launch
• NVIDIA developer blog — Marble x Isaac Sim
• Microsoft Research — Introducing Muse
• Microsoft Research — WHAM publication
• Microsoft Research — WHAMM real-time modelling
• Microsoft — WHAM on HuggingFace
• Meta AI — Code World Model (CWM) research
• Meta Quest Blog — WorldGen research update
• arXiv 2511.16825 — WorldGen: From Text to Traversable and Interactive 3D Worlds
• Hacker News — WorldGen discussion
• Promethean AI — official site
• Promethean AI — Unreal plugin, GitHub
• Promethean AI — GitHub org
• Tencent-Hunyuan — HY-World 2.0, GitHub
• Tencent — HY-World 2.0, HuggingFace
• Tencent — HY-World 2.0 technical report PDF
• ToKnow.ai — HY-World 2.0 coverage
• CnTechPost — Tencent open-sources HY-World 2.0
• BuildFastWithAI — Tencent & Alibaba world models comparison
• HuggingFace — HunyuanVideo LICENSE (territorial-restriction pattern)
• DeepWiki — HunyuanWorld-Mirror commercial licensing
• GitHub — Hunyuan3D-2 LICENSE
• innFactory — Tencent Hunyuan (Hy3) generation, Apache 2.0 note
• Bloomberg — Alibaba AI model for gaming development (Happy Oyster)
• 36Kr — ByteDance's 2026 AI priorities
• ChinaBizInsider — ByteDance AI strategy 2026
• Runway Research — Introducing GWM-1
• TechCrunch — Runway's first world model
• NVIDIA — Cosmos product page
• NVIDIA developer blog — Cosmos 3 physical AI
• NVIDIA blog — How Cosmos 3 helps physical AI think before it acts
• HuggingFace blog — NVIDIA Cosmos 3 launch
• MarkTechPost — NVIDIA Cosmos 3 release
• HuggingFace — nvidia/Cosmos3-Nano
• Spheron Blog — deploying Cosmos 3 on GPU cloud
• WaveSpeed — What is NVIDIA Cosmos3-Nano
• Niantic Spatial — Large Geospatial Model / world models 2026
• Niantic Labs — Building a Large Geospatial Model
• MIT Technology Review — Niantic geospatial data training robots
• CSM AI — official site
• The Motley Fool — Unity CEO on world models vs. game engines
• Nasdaq — same Unity CEO coverage
• iconpolls.com — Unity AI review 2026, Muse deprecation
• Unreal Engine 5.8 — official release notes
• Unreal Engine — Procedural Content Generation overview, 5.8 docs
• Sean Kim — UE 5.7 PCG framework at GDC 2026
• Creative Bloq — UE 5.7 PCG update explained
• StraySpark — AI-assisted procedural level design via MCP
• CG Channel — NanoGS free Gaussian Splatting plugin for UE5
• Splatware — Gaussian Splatting in UE5, capture-to-game pipeline 2026
• Epic Developer Community — Gaussian Splatting options forum thread
• Engadget — Epic Games embracing generative AI, UE6 details
• GIGAZINE — Unreal Engine 6 early look, Gemini/Claude integration
• PlayDay.one — Unreal Engine 6 targets late 2027 Early Access
• BuildFastWithAI — World Models race 2026 comparison

---

Relationship to the existing pipeline research (no contradiction, one addition)

Nothing here changes RESEARCHED_STACK.md Chain 2's ratified world chain (Copernicus DEM → GDAL

alteration → Gaea/UE Mesh Terrain detail synthesis → PCG Biome Core populate). It adds one concrete

recommendation: when the region-page template v2 lands (Phase C), the PCG-DataTable-ingestion path is

confirmed as the right target for zone/site rows, and no world-model or scene-assembly product

surveyed here should be adopted as a replacement for it — at most, Marble or HY-World 2.0 earn a

narrow, clearly-labeled role as disposable previz reference, gated behind the same schema-bound

discipline as everything else this project builds.

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