COSMOS3_DEEP_DIVE.md

pipelines/COSMOS3_DEEP_DIVE.md

NVIDIA Cosmos 3 — Deep Dive for the Humanity Pipeline

Status: research brief, not canon. Scoped specifically to Cosmos 3 (the omnimodel), requested

separately from NVIDIA_WORLDGEN.md (same folder, same research date) which answers a narrower

question — "does an NVIDIA product replace the Cesium/Gaea2Unreal/SRTM terrain chain." This brief

re-derives Cosmos 3 specifically, from scratch, against primary sources, per the request. Research

date 2026-07-15. Every substantive claim is tagged VERIFIED (checked against a primary or

directly-fetched source, URL inline) or INFERRED (my own reasoning bridging a gap the sources

didn't state directly, or a secondary-source synthesis not independently cross-checked). Where

sources conflicted or a claim mattered but couldn't be pinned down, that's stated explicitly rather

than smoothed over.

---

0. Premise check — read this first

**"NVIDIA Cosmos 3" is real. It shipped. The specific framing in the research brief ("world's first

fully open omnimodel," "Mixture-of-Transformers," "first omni-model with native reasoning, world and

action generation") is accurate, not marketing exaggeration invented upstream.** This was not

assumed — it was checked cold against NVIDIA's own newsroom, investor relations, developer blog,

Hugging Face, and the arXiv technical report, cross-referenced against independent tech press.

VERIFIED, multiply-sourced:

Taipei / COMPUTEX 2026 by Jensen Huang.

(NVIDIA Newsroom;

NVIDIA Investor Relations;

NVIDIA Cosmos product page)

across text, image, video, ambient sound and action" is NVIDIA's own stated language, not a

third-party paraphrase.

(nvidia.com/en-us/ai/cosmos)

with a diffusion (DM) generation tower via joint attention, not a rebrand of a diffusion-only or

MoE model.

(Cosmos 3 technical report, arXiv 2606.02800;

NVIDIA Developer Blog;

Hugging Face launch blog)

(a Colab tutorial built on Cosmos 3), consistent with a mid-July 2026 "today" — the premise's

internal dating holds together and required no correction.

One correction to the task's own framing: the prompt assumed SIGGRAPH 2026 falls in August.

It doesn't — SIGGRAPH 2026 runs July 19–23, 2026, in Los Angeles, i.e. within days of this

research being written, not weeks away.

(nvidia.com/en-us/events/siggraph;

s2026.siggraph.org). See §5.

Relationship to NVIDIA_WORLDGEN.md (sibling doc, same folder, same day): that brief's verdict —

"Cosmos outputs video/images/action, not meshes or USD; it cannot be a terrain or asset generator no

matter how you wire it" — is independently reconfirmed by this deeper, Cosmos-3-specific pass (see

§1, §3). Two things are worth flagging back to that document, not as a contradiction but as a

refinement:

1. Its Cosmos VRAM/model-size row (§1.1: "2B variant... 14B variant...") describes

Cosmos-Predict2.5's model matrix, not Cosmos 3's. Cosmos 3 proper ships as Nano (16B) and

Super (64B), with an Edge (2B) variant announced but not yet released — see §1 below for the

corrected figures, directly sourced from the Cosmos 3 launch materials rather than the prior

generation.

2. That brief's scope question was narrower ("does this replace the terrain-authoring chain") and

didn't investigate the broader NVIDIA world-model research orbit. This brief did, and found real

(if licensing-encumbered) bridge candidates — Lyra 2.0 and SANA-WM — that are worth the pipeline

review's attention even though neither is "Cosmos" and neither replaces Cesium/Gaea/SRTM either.

See §3.

---

1. What Cosmos 3 actually is and ships

Architecture

Two-tower Mixture-of-Transformers (MoT): a Reasoner (autoregressive transformer, vision-language

model — interprets text/image/video input, does spatial grounding, temporal reasoning, question

answering, chain-of-thought about physical scenes) fused via joint attention with a Generator

(diffusion transformer — produces physics-aware video, images, audio, and action outputs conditioned

on the Reasoner's understanding). Separate parameter sets for the AR and DM subsequences, unified

attention space. A single model does both understanding and generation, eliminating the

orchestration NVIDIA previously needed across separate Cosmos Predict / Transfer / Reason models (see

§4). VERIFIED

NVIDIA Developer Blog,

Hugging Face blog,

technical report.

A known architectural quirk, self-disclosed in the technical report: without care, the last language

token and the first frame's vision tokens land at nearly identical temporal positions, causing

over-saturation and checkerboard artifacts in early generated frames; NVIDIA's fix is a fixed temporal

gap inserted between the AR and DM subsequences. VERIFIED

technical report PDF.

Model family and sizes (corrected vs. the prior-generation figures)

VariantTotal paramsSplitTarget hardwareStatus
Cosmos 3 Nano16B8B Reasoner + 8B GeneratorWorkstation (RTX PRO 6000 class; also the one that fits, tightly, on a single RTX 5090 — see below)Shipped 2026-05-31
Cosmos 3 Super64B32B Reasoner + 32B GeneratorDatacenter (Hopper/Blackwell, e.g. GB200)Shipped 2026-05-31
Cosmos 3 Edge2BReal-time edge inferenceAnnounced, not yet released

Specialized checkpoints also shipped on top of Super: Cosmos3-Super-Text2Image and

Cosmos3-Super-Image2Video (both 64B); a robot-control policy checkpoint,

Cosmos3-Nano-Policy-DROID (16B), also shipped. VERIFIED

NVIDIA Newsroom,

developer.nvidia.com blog,

nvidia/Cosmos3-Nano model card,

GitHub — nvidia/cosmos.

Inputs / outputs — concrete formats (from the Cosmos3-Nano and Cosmos3-Super-Image2Video model

cards, VERIFIED, huggingface.co/nvidia/Cosmos3-Nano,

huggingface.co/nvidia/Cosmos3-Super-Image2Video):

reasoning-only tasks.

(8-bit sRGB) — grayscale explicitly unsupported.

not a long input video).

robot/vehicle control-signal format (joint angles, steering, throttle), **not a generic 6-DoF

cinematic camera path**. That distinction matters for §2/§3 below.

resolution tiers 256p/480p/720p; a documented alternate spec allows 1–8 seconds by duration. 720p is

the observed ceiling — no 1080p/4K output confirmed anywhere in this research pass.

out:** strings.

format specification, anywhere.** Output is pixels, audio samples, text tokens, and control-signal

numbers — never geometry.

License

OpenMDW-1.1, a Linux Foundation license NVIDIA newly adopted across the Cosmos, Isaac GR00T, Ising,

and Nemotron families. **This is a genuinely permissive, model-specific open license — not the older,

more restrictive NVIDIA Open Model License**, and it is worth being precise about since a lot of

secondary coverage conflates the two. Fetched and read directly from the license text itself

(VERIFIED, raw license text via OpenMDW/OpenMDW GitHub,

Linux Foundation press release):

copyright, patent, database, and trade secret rights" — use, modify, redistribute, train derivative

models, all permitted.

— i.e. no claim over what you generate with it.

Model Materials, your own rights under the license terminate (standard Apache-style defensive

clause).

AI or excludes entertainment/games. (The *model card's* stated "intended use" is physical AI — see

below — but that's a usage recommendation, not a license carve-out.)

deployment geography, no MAU cap, no excluded territories — a materially cleaner license posture

than, e.g., the Hunyuan3D Community License covered in the sibling LOCAL_3D_ASSET_GEN.md doc (MAU

cap + territory exclusions).

Hardware / VRAM — the RTX 5090 32GB question, answered directly

This is the single most concrete, load-bearing fact for a solo dev's incoming RTX 5090, so it's worth

citing precisely rather than paraphrasing. A hands-on install/run report, dated 2026-06-04 (four days

after launch), documents actually running Cosmos 3 Nano on a real RTX 5090 (32GB):

precision (the model card states FP4/FP8/FP16 are "not officially supported" for the raw

Diffusers/Transformers checkpoint). That leaves under 3GB of headroom on a 32GB card for

everything else (VAE, text encoder, activations, KV cache), which is not enough on its own.

PyPI release didn't yet support the Cosmos3OmniPipeline), setting

PYTORCH_CUDA_ALLOC_CONF before importing torch to fight memory fragmentation, and switching from

component-level offload to enable_sequential_cpu_offload() to avoid outright VRAM exhaustion.

quality was rated by the author as having "a certain atmosphere... but the fine details definitely

need work" — usable for local experimentation, not production-ready without further prompt tuning.

VERIFIEDRunning NVIDIA Cosmos 3 Nano on an RTX 5090.

distribution, distinct from the raw HF checkpoint) supports **BF16, FP8, or NVFP4 quantized

checkpoints, with NVFP4 giving up to 2x inference speedup. VERIFIED** —

developer.nvidia.com blog.

Whether a quantized NIM checkpoint is actually deployable on a consumer RTX 5090 (vs. the RTX PRO /

datacenter targets NIM usually assumes) **was not confirmed either way in this research pass —

flag as open/INFERRED-plausible, not verified.** If true, it would meaningfully ease the tight

32GB fit described above.

consumer-card target at all. VERIFIED

developer.nvidia.com blog.

("Looking forward to trying this out on my $10000+ workstation"), pricing the *meaningfully usable*

tier (not just "technically boots") above consumer reach. That's a sentiment data point, not a hard

spec, but it corroborates the "tight, not comfortable" read on the Nano/RTX 5090 fit.

**Bottom line on hardware: Cosmos 3 Nano runs on one RTX 5090 (32GB), but at the edge of the card's

memory, in BF16 only, requiring offload workarounds a stock install doesn't provide out of the box,

and at 720p/sub-8-second clip lengths. It is not a comfortable local workflow yet.**

Training data and benchmarks

Trained on 20 trillion tokens of multimodal data — nearly a billion images, 400 million real and

synthetic videos, ambient audio, text, and action data from humans and robots. VERIFIED

search-aggregated from NVIDIA sources, corroborated across multiple outlets.

Post-trained Cosmos 3 models rank #1 among open models for text-to-image and image-to-video

(Artificial Analysis) and #1 open robot policy model (RoboArena) at the time of the technical report.

VERIFIEDtechnical report,

Hugging Face blog. Note this is a

same-generation open-model comparison, not a claim of beating closed frontier models — the Hacker News

discussion (§2) explicitly notes it currently underperforms closed models like Grok Imagine and

Seedance2 on general video quality.

Stated intended use (from the model card itself)

"Physical AI: Encompassing robotics, autonomous vehicles (AV), and smart space environments, including

industrial and factory-scale applications." The card explicitly warns outputs should not inform

"robotics control, autonomous systems, scientific simulation, or safety-critical planning" without

additional validation, and that the model is not suitable for "physically accurate simulation,

reliable ground-truth reasoning, or safety-certified decision making." VERIFIED

nvidia/Cosmos3-Nano model card. This is a usage

recommendation/disclaimer, not a license restriction (see License above) — nothing stops non-physical-

AI use contractually, but NVIDIA is explicit that the model wasn't built or validated for it.

---

2. Game-dev applicability — argued both directions

The case FOR

signal that NVIDIA sees crossover value even though Cosmos 3's own positioning is physical AI.

Founding members: Agile Robots and Skild AI (robotics), Generalist (embodied AI), and — notably —

Black Forest Labs (makers of the FLUX image-generation models), Runway (AI video generation

for film/creative production), and LTX (Lightricks' LTX-Video, a creative video-gen model). Three

of six founding partners are creative-AI companies, not robotics companies. VERIFIED

NVIDIA Newsroom.

What exactly they'll build together was not detailed in any source found — treat the Coalition as a

signal of intent, not a shipped capability.

itself.** Cosmos Transfer models (the specialist companion line, see §4) take spatial control inputs

including depth, segmentation, edges, and high-definition maps to condition photorealistic

generation from ground-truth 3D scenes. HD-map-style conditioning is the closest existing Cosmos

capability to "real geodata → generated environment," even though it's built for AV synthetic-data

augmentation, not region art-direction. VERIFIED

SIGGRAPH 2026 coverage of Cosmos Transfer-2 (see

§5), corroborated by the Omniverse SDG reference workflow docs.

for region art direction (Cosmos3-Super-Text2Image, Cosmos3-Super-Image2Video) — this is a real,

shipped capability, though nothing in the evidence suggests Cosmos 3 is competitive with dedicated

creative image/video models for this purpose (it isn't benchmarked against them, and its training

data was "curated for physical relevance," per NVIDIA's own description, not for aesthetic/creative

range). INFERRED as technically possible, not verified as a good idea relative to alternatives.

underlying Cosmos technology "could be tuned" for entertainment, VR, and "immersive game

environments." **This is aspirational framing from a secondary source, explicitly tagged INFERRED —

it describes potential, not a shipped or demonstrated capability.** —

resources.nvidia.com marketing collateral.

The case AGAINST

takes one image plus a text prompt and generates a coherent video continuation — it does not

accept a camera trajectory and is not a fly-through generator. The "action" modality in Cosmos 3 is

robot/vehicle control-signal space (joint angles, steering), not cinematic camera control. VERIFIED

nvidia/Cosmos3-Super-Image2Video model card.

This directly contradicts the FOR-case hope of "camera-controlled world exploration" — that

capability exists in NVIDIA's world-model portfolio (§3), just not in Cosmos 3.

model cards, the GitHub repo, and the technical report. Output is video/image/audio/action/text,

full stop.

the launch: commenters immediately clarified Cosmos 3 is a robotics/AV world model, not a

competitor to consumer creative tools (Nano Banana, Seedance2); pricing sarcasm about the

"$10,000+ workstation" needed for meaningful use; criticism that demo videos showed pedestrians not

reacting to traffic, vehicles running red lights, and unrealistic shadow physics — i.e. the physical

plausibility Cosmos is *for* was itself questioned; a technical critique that the two-tower

architecture "bakes in" too much prior structure and may plateau against scaling-law competitors.

**No mention of game development, creative, or non-robotics use appeared anywhere in that

discussion. VERIFIED** — Hacker News, item 48356654.

checkerboard artifacts in early frames (architectural, mitigated but not eliminated); "can still

produce artifacts in long, high-resolution, or physically complex outputs"; a named failure mode

of "over-trusting plausible video" — a clip can look physically reasonable while violating sensor

geometry or real-world physics. VERIFIED

technical report.

miniature/reimplementation of the Cosmos 3 architecture for learning purposes (MarkTechPost,

2026-07-08); the RTX 5090 install/feasibility report cited above; robotics/AV-focused synthetic-data

tutorials and NIM deployment guides. **No game-development-specific project, demo, or experiment

using Cosmos 3 was found anywhere in this research pass**, despite specifically searching for indie/

game-dev hands-on reports, Reddit discussion, and environment-art/procedural-generation use. That

absence is itself informative for a launch that's six weeks old with a genuinely open, commercially-

usable license — if there were an obvious game-dev angle, the permissive license removes the usual

barrier to someone having already tried it publicly.

Net read on Q2: the FOR case is real but structural/adjacent (the Coalition's creative-company

membership, Transfer's geodata-conditioning capability) rather than something Cosmos 3 itself does for

a game today. The AGAINST case is direct, model-card-confirmed, and corroborated by actual developer

reaction. Cosmos 3 itself is not close to a game-dev tool; the more interesting FOR-case material lives

one step outside Cosmos 3 proper (§3).

---

3. The bridge question — Cosmos-family output to a usable Unreal asset

Cosmos 3 itself has no bridge. No NVIDIA-provided exporter, no USD hook, no depth-to-geometry path

was found for Cosmos 3's own outputs. Anything downstream would have to start from raw video/image

frames and go through a completely generic (non-NVIDIA) video-to-3D reconstruction pipeline, with all

the consistency risk that implies (see below).

But the research surfaced two adjacent NVIDIA research projects — not Cosmos, not on the Cosmos

Lab research page, but from NVIDIA's other world-model labs — that are much closer fits to what a

"bridge" would need to look like. Getting the Cosmos/non-Cosmos distinction right matters, because

secondary coverage sometimes blurs them together under a loose "NVIDIA world models" umbrella.

Lyra 2.0 — the closest fit, and the catch

From NVIDIA's Spatial Intelligence Lab (research.nvidia.com/labs/sil/, distinct from

research.nvidia.com/labs/cosmos-lab/) — not a Cosmos product.

prompt), generates a long camera-controlled walkthrough video, then reconstructs that video into

explicit 3D Gaussian Splats and surface meshes via a feed-forward 3DGS pipeline fine-tuned to

tolerate generative-model inconsistency, plus a hierarchical sparse-grid mesh extraction step.

Exports .ply splats and meshes intended for real-time engines and physics simulators — the

project page names Unreal, Unity, and Blender explicitly as target pipelines. VERIFIED

Lyra 2.0 project page,

arXiv 2604.13036.

would otherwise sink a "generate then reconstruct" pipeline — spatial forgetting (maintains

per-frame 3D geometry to retrieve and warp relevant past frames) and temporal drifting

(self-augmented training on degraded outputs, teaching the model to correct its own drift). This is

a genuine, purpose-built answer to the "AI video is too inconsistent for 3D reconstruction" problem

— not a naive video-then-photogrammetry bolt-on.

published 2026-04-15; GitHub repo (nv-tlabs/lyra) has 2.2k stars and active development. Source

code is Apache 2.0. VERIFIED

github.com/nv-tlabs/lyra.

under the "NVIDIA Internal Scientific Research and Development Model License," which is explicitly

non-commercial.** Verbatim: *"The Model and any Derivative Model may not be distributed, deployed,

sublicensed, publicly displayed, publicly performed, or sublicensed. You may not use the Model or a

Derivative Model in a production environment or for the purpose of generating works for sale or

distribution."* VERIFIED

huggingface.co/nvidia/Lyra-2.0. Code (Apache 2.0) and

weights (non-commercial research license) are under different, inconsistent licenses — a

meaningful trap for anyone who reads "open source, Apache 2.0" in secondary coverage (some

secondary sources reported it that way) and assumes the whole package is commercially usable. It

is not, as shipped. Custom/commercial licensing is available through NVIDIA's research-licensing

channel but wasn't priced or scoped in anything found here.

SANA-WM — commercially clean, but video-only

From NVLabs (the SANA image-model lineage) — again not a Cosmos product, despite being

topically adjacent (world modeling) and released within weeks of Cosmos 3.

trajectory, generating a spatially consistent 60-second, 720p video following that path.

VERIFIEDarXiv 2605.15178,

MarkTechPost coverage.

a 60-second clip in 34 seconds (2.1x real-time). This is a materially more comfortable

consumer-GPU story than Cosmos 3 Nano's tight 29GB+ fit. VERIFIED — same sources above.

(NVlabs/Sana). Genuinely commercially usable, no research-only trap. VERIFIED** —

github.com/NVlabs/Sana.

describes SANA-WM producing splats, meshes, or any 3D representation — output is an MP4. To become

a usable 3D asset, its video output would need to go through a *separate*, non-NVIDIA reconstruction

step (below).

The generic reconstruction-and-import toolchain (fully independent of NVIDIA)

This is the layer that would have to sit between a generated video (from SANA-WM, or in principle from

Cosmos 3's own image-to-video output) and a usable Unreal asset, since neither Cosmos 3 nor SANA-WM

does 3D reconstruction natively:

independent of anything NVIDIA shipped for this purpose — Postshot (Jawset, Windows desktop,

fully local NeRF+3DGS reconstruction, no cloud dependency), Polycam and Luma AI (mobile/web,

more turnkey), Nerfstudio and KIRI Engine as additional options. Typical export: .ply

splats, with .glb/.usdz mesh fallbacks (Luma generates an approximate mesh alongside the splat).

VERIFIED, general 2026 tooling landscape —

thefuture3d.com tool comparison,

Postshot review.

Unreal Engine as of this research** — one source notes Epic's roadmap shows interest but "5.8 may or

may not change the picture," i.e. unresolved at research time; this should be verified directly

against Epic's UE 5.8 release notes before relying on it. A mature third-party plugin ecosystem

fills the gap regardless: NanoGS (Nanite-style LOD/culling for splats, confirmed for UE 5.6/5.7),

the Luma AI Unreal plugin (described by one source as "the production-relevant first-party path in

2026"), plus several community plugins (MLSLabsGaussianSplattingRenderer-UE, XScene-UEPlugin,

unreal-splat, SplatRenderer). 60fps at sub-1M-Gaussian scene sizes is reported as achievable on

mid-range hardware. VERIFIED

Gaussian Splatting in Unreal Engine 5: capture-to-game pipeline 2026,

CG Channel on NanoGS.

Assembling the hypothetical chain — and its real risk

A full chain from "region prompt" to "explorable Unreal environment" would look like: SANA-WM

(commercial, RTX-5090-capable, camera-controlled video) → Postshot or Nerfstudio (commercial,

mature, video-to-3DGS/mesh) → a third-party splat plugin (NanoGS / Luma AI plugin / etc.) → UE 5.8

scene. **Every individual link in that chain is independently verified to exist and be commercially

usable.** No one has been found, in this research pass, actually running that specific chain

end-to-end — this exact combination is INFERRED as plausible, not verified as demonstrated.

The real technical risk, grounded in evidence rather than speculation: general-purpose 3DGS

reconstruction tools (Postshot, Nerfstudio, etc.) were built and tuned for **real, multi-view-

consistent photogrammetry** captures — many overlapping photos or a real camera's video of a real,

static scene. Generative video, even camera-conditioned generative video, produces frames that are

each individually plausible but not necessarily *geometrically* consistent frame-to-frame the way a

real capture is — this is precisely the "spatial forgetting" and "temporal drifting" problem Lyra 2.0

was purpose-engineered to solve (§3 above), and precisely the kind of physical/geometric-plausibility

gap the Hacker News critique and Cosmos's own technical report flagged for Cosmos 3's video output

(§2). Feeding SANA-WM's output into a reconstruction tool that assumes real-capture-grade consistency

could plausibly produce splat artifacts, floaters, or warped geometry that Lyra 2.0's specialized

pipeline was specifically built to avoid. **This is reasoned inference from the technical evidence

gathered this session, not a tested result — it's the reason this chain should be bench-tested in

a small pilot before any pipeline dependency is built on it**, not assumed to work at Lyra-2.0 quality

just because the individual pieces are commercially licensed.

**This section is a genuine complement to NVIDIA_WORLDGEN.md §1.5's TRELLIS-NIM Blueprint finding,

not a duplicate of it.** That Blueprint generates discrete objects/props (up to ~20 per prompt,

object-level GLBs) and is already confirmed shipped, local, and working. What this section covers is

environment/world-scale reconstruction from a generated fly-through — a different layer of the

pipeline (region/scene backdrop vs. individual props), less mature, and licensing-encumbered at its

best-engineered option (Lyra 2.0).

---

4. Cosmos's place in its own ecosystem

The Predict / Transfer / Reason lineage

Cosmos 3 consolidates, rather than simply sitting alongside, the previous generation of

specialist Cosmos models:

directly into Cosmos 3 as its Reasoner tower — it is no longer a separate model to deploy.

VERIFIEDdeveloper.nvidia.com blog.

specialist model, released October 2025) is now explicitly in limited-maintenance mode — NVIDIA

states the repo "is no longer under active development" and directs users to Cosmos 3 for current

capability. VERIFIED

github.com/nvidia-cosmos/cosmos-predict2.5.

(This is the model whose 2B/14B figures appear in NVIDIA_WORLDGEN.md §1.1 — accurate for

Predict2.5, not for Cosmos 3.)

depth/segmentation/edge/HD-map inputs" specialist, also October 2025) **continues to exist as a

separate, actively-referenced companion model** rather than being fully absorbed into Cosmos 3 —

it's explicitly used alongside Cosmos 3 in NVIDIA's own SIGGRAPH 2026 hands-on lab materials (§5).

VERIFIEDnvidia.com/en-us/events/siggraph.

A further evolution, referred to in some coverage as "Cosmos Transfer-2" ("coming soon,"

simplified prompting, 1-step distillation replacing a 70-step process, targeting RTX PRO Servers),

was reported for SIGGRAPH 2026 — but the primary NVIDIA event page names the SIGGRAPH hands-on lab

model as "Cosmos-Transfer2.5," not "Transfer-2." **The exact naming/version relationship between

"Transfer 2.5" and "Transfer-2" is unresolved in the sources gathered this session — flag as an

open naming ambiguity rather than asserting a clean 1 → 2.5 → 2 progression.**

NIM microservice availability

and tensor parallelism, plus Efficient Video Sampling (EVS) for token reduction.

This means the optimized, quantization-supporting deployment path (§1's FP8/NVFP4 note) is currently

half-available at best — the generation side a solo dev would actually want isn't in NIM form yet;

today's local option is the raw Diffusers/Transformers checkpoint (the BF16, 29GB+, offload-required

path from §1).

Omniverse / USD relationship

Cosmos pairs with Omniverse as a synthetic-data multiplication loop, not a content-authoring

pipeline: Omniverse/USD builds a 3D scenario, a Replicator writer (CosmosWriter) exports depth/

segmentation/canny AOVs as conditioning video, and Cosmos Transfer generates photorealistic variations

from those control signals — closing the domain gap between clean synthetic renders and real-world

messiness for robot/AV training. VERIFIED

Omniverse SDG reference workflow,

Into the Omniverse: OpenUSD Workflows for Physical AI.

This is the same USD/Omniverse plumbing NVIDIA_WORLDGEN.md covers for Cesium-for-Omniverse — worth

noting it's shared infrastructure, but the Cosmos-specific use of it (robotics/AV synthetic data) does

not open a new game-content path beyond what that sibling doc already concluded about Omniverse itself

(useful as an optional USD staging layer, not a generator).

---

5. Roadmap — what's next, and the SIGGRAPH 2026 correction

SIGGRAPH 2026 is July 19–23, 2026, at the LA Convention Center — not August. This is close enough

to the research date (2026-07-15) that the keynote (Monday, July 20, 3:45–4:45pm PT — "Next Era of

Graphics: Neural Rendering, World Models, and Simulation," presented by NVIDIA Research/Engineering

leads Neil Ashton, Ming-Yu Liu, and Edward Liu) is days away, not weeks. VERIFIED, first-party

nvidia.com/en-us/events/siggraph,

s2026.siggraph.org keynote page.

What's confirmed for SIGGRAPH 2026:

July 20, 1:30–3pm PT), walking through Cosmos 3 Nano Reasoner (video captioning, augmentation

prompting, chain-of-thought validation) chained with Cosmos 3 and Cosmos-Transfer2.5 (photorealistic

synthesis) and Cosmos 3 action modes (translating scene understanding into embodied control

commands). VERIFIED, first-party. Entirely robotics/physical-AI framed — no game-engine or game-

development content in the session description.

§4) targeting simplified prompting, faster photoreal synthetic data generation, and RTX PRO Server

deployment.

Epic Games or Unity partnership with NVIDIA turned up nothing — the one Epic/Unity partnership news

found in 2026 (Unity games publishable into Fortnite, Unity commerce tools coming to Unreal) is

confirmed unrelated to NVIDIA or Cosmos entirely, a red herring worth ruling out explicitly

rather than leaving ambiguous.

language is the aspirational third-party marketing copy already flagged as INFERRED in §2 — nothing

resembling a committed roadmap item.

---

CAPABILITY BOUNDARY

What follows is a strict read of the evidence gathered this session — not the hype framing this

research request opened with, and not a rubber-stamp of the old "robotics only, no meshes" dismissal

either. Both get checked against what was actually found.

What Cosmos 3 (the real, current version) can contribute to this UE 5.8 pipeline TODAY

Almost nothing, directly. The honest list:

concept-art aid for region art direction (Cosmos3-Super-Text2Image / Image2Video) — technically

possible, not verified to be competitive with dedicated creative-image tools the project may already

have access to, and NVIDIA's own training-data description ("curated for physical relevance") cuts

against it being a strong creative tool. Low value, not zero.

understanding Mixture-of-Transformers world models, of interest to the "generative-entity-system-

vision" research thread but not a pipeline component.

region/geodata ingestion beyond what Cosmos Transfer (a different, companion model) does for

robotics synthetic data.

What it could contribute WITH bridging tools

A real, evidence-grounded (but unproven-as-a-chain) path exists — and it does not run through Cosmos 3

at all:

Postshot or Nerfstudio (commercial, mature, local, video/photo-to-3DGS-and-mesh) → a third-party

Unreal splat plugin (NanoGS, the Luma AI Unreal plugin, or equivalents) → an explorable 3DGS

environment in UE 5.8. Every link is individually real and commercially usable. The chain as a whole

has not been demonstrated by anyone found in this research and carries a real, evidence-grounded risk

(generative-video-to-photogrammetry-reconstruction consistency) that should be pilot-tested on a

small scene before any dependency is built on it. Best case, this is a **previz/blockout/reference

environment generator**, not a source of shippable game geometry.

purpose-built engineering against exactly the consistency risk flagged above — the best-engineered

option found in this research. But its model weights are licensed non-commercial-research-only

(verbatim restriction quoted in §3): no production use, no distribution, no generating works for sale.

Usable today for personal evaluation/prototyping/technical spikes only. Turning it into something

shippable requires either a custom commercial license from NVIDIA (unpriced/unscoped in anything

found here) or retraining from the Apache-2.0 source code — a serious compute and data undertaking,

almost certainly not worth it for a solo developer relative to the mature independent tools above.

Cosmos family with real geodata-adjacent relevance, but it's built and evidenced for robotics/AV

synthetic-data augmentation, not region art-direction — treating it as a bridging tool for this

pipeline would be speculative beyond what the evidence supports.

What it cannot contribute at all

NVIDIA_WORLDGEN.md's verdict on the same point.

DEM; Cesium remains the actual geodata leg exactly as that sibling doc concluded.

exists; the Omniverse relationship that does exist (§4) is a synthetic-data-generation loop for

robotics, not an asset-authoring bridge.

just to generate a single short offline clip), wrong license posture for the closest real bridge tool

(Lyra 2.0), and wrong output type (video/action tensors, not renderable geometry) even before

runtime performance is considered.

Overall verdict

The old dismissal — "robotics infrastructure, no meshes" — **holds up, technically, under this fresh,

thorough, Cosmos-3-specific re-examination.** That was not assumed going in, and it would have been

reported the other way if the evidence had pointed there; it didn't. Where the old dismissal falls

short is in stopping at Cosmos: the more interesting, genuinely-new finding this deeper pass surfaced

is that **NVIDIA's adjacent world-model research (Lyra 2.0, SANA-WM) — released within weeks of Cosmos

3, easily confused with it, but organizationally and technically distinct — contains a real, if

currently either licensing-blocked or reconstruction-risk-laden, path from "region prompt" to

"explorable 3D environment."** That path is worth a small, deliberate bench-test once the RTX 5090

workstation lands (SANA-WM → Postshot → a UE splat plugin, on one test region), scoped explicitly as a

previz/reference experiment, not adopted as a production pipeline stage on the strength of this

research alone.

---

Sources

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