Choose a 3D world workflow

Compare explorable 3D worlds, measured Gaussian scenes, procedural geometry, and action-conditioned video.

What kind of world do you need?

How the data moves
  1. InputChoose the scene inputprompt · photograph · calibrated views · random seed

    Different inputs lead to different kinds of uncertainty.

  2. ProcessKeep the representation explicit3DGS · mesh/materials · collider · RGB video

    A video is not a reusable 3D scene, and a splat is not a collision surface.

  3. OutputRender, inspect and simulateSpark browser view · Blender · CARLA ego camera

    Use our recordings and code; check model terms before running a generative checkpoint.

A camera-controlled video and an explorable 3D scene are different outputs. Start by deciding whether you need new pixels, a persistent radiance field, editable geometry, or a world with collision and dynamics.

Input and method Output What the next application can do
Text/image → generative 3D model Gaussian scene; sometimes a mesh or collider Render unseen viewpoints; import supported representations into another application
Calibrated photos → 3DGS reconstruction Centers, scales, rotations, opacity and color Re-render a measured space near captured views
Seed + procedural rules → scene generator Meshes, materials, lights and cameras Edit geometry, export assets, add physics and create datasets
State + controls → driving simulator Next state, sensor frames and event records Test controls in a closed loop and record ego views
Image + camera actions → video model A time sequence of RGB images Inspect generated visual motion; a reusable 3D asset is not implied

Generate a world yourself: WorldGen text/image → 3D scene. Our working path: CARLA ego-camera simulation → known views → fVDB 3DGS → interactive Spark viewer. A separate Infinigen house tour shows actual generated geometry. Cosmos video generation remains available as a different output type.

Compare the tools and their data boundaries

Generative 3D worlds

World Labs Marble accepts prompts and visual inputs to produce explorable worlds. Its export documentation describes SPZ/PLY splats and separate mesh/collider options. Hosted generation is not a local open-source model. Free generation and export rights are separate; current exports require paid plans. We did not call a paid API or generate a Marble world in this run. Spark, the open-source World Labs renderer, is what our site uses to display our own fVDB result. Marble export · File specifications · Spark source / MIT

NVIDIA Lyra 1.0 combines generated multiview video information and a Gaussian decoder to output 3D/4D scenes. Its model card lists the NVIDIA Open Model License. Lyra 2.0 adds long-horizon, 3D-consistent world exploration and reconstruction; its model weights instead use the NVIDIA Internal Scientific Research and Development Model License. Both implementations are public Apache-2.0 code. Do not transfer one version's weight license to the other. The official Lyra 2 setup was tested on H100 hardware; neither version was executed in this set of experiments. Lyra source · Lyra 1 model card · Lyra 2 code and model terms

WorldGen generates a panorama from text or an image, predicts distance with DA-2, and converts it into splats and a mesh. After both FLUX approvals, we executed two text scenes and one calibrated-image scene on an RTX workstation. The hands-on lab includes actual 20-second free-viewpoint recordings, click-to-load 3D previews, source code and geometry checks. Code and LoRA are Apache-2.0; FLUX weights retain separate noncommercial model terms and user-specific access gates. Our image route uses a documented known-camera projection adapter. WorldGen · FLUX.1-dev · FLUX.1-Fill-dev

Measured reconstruction and editable geometry

fVDB Reality Capture optimizes a radiance field from posed images. In our run, known CARLA cameras and training-view depth seed the Gaussians; there is no prompt-driven scene invention. SHARP infers nearby-view Gaussians from one image, with greater uncertainty in hidden regions. SHARP's code uses Apple's custom software license and its weights permit noncommercial research. Infinigen creates meshes, materials and room layouts with procedural rules. It is not a neural foundation model. Its BSD-3-Clause code and the Blender/asset dependencies have separate scopes. fVDB · SHARP licenses · Infinigen

Run status and practical limits

Verified on 2026-11-20. “Executed” means our own run and saved artifacts, not a replay of the author's promotional video.

Tool Execution status Available result
CARLA 0.9.16 / UE4.26 Executed on RTX 5090 Four ego-camera environments, chase originals and trajectories
fVDB Reality Capture 0.6.0 Executed on RTX 5090 72-view dataset, full PLY, 3DGS video, 6.15 MB browser preview
Apple SHARP Executed on RTX 5090 Two single-image reconstructions and nearby-view recordings
Infinigen / Blender Scene generation and CPU Cycles render Furnished house, saved .blend, camera path and tour recording
Spark 2.4.0 Integrated into the site Click-to-load WebGL2 viewer; rendering uses the visitor's device
WorldGen / DA-2 Executed on RTX 5090 with FLUX FP4/offload 3 generated worlds, Gaussian PLY + mesh, translated camera recordings and 3 browser previews
NVIDIA Lyra 1 / 2 Official sources and licenses checked Reference workflow; not executed in this update
World Labs Marble Hosted reference No account, payment or API job started

WonderWorld's official installation calls for 48 GB GPU memory, above the tested 32 GB GPU. HunyuanWorld 1.0's community license excludes South Korea, so it was not selected for this run. These are concrete selection limits, not a comparison of visual quality. WonderWorld · HunyuanWorld 1.0 license

  1. Drive in CARLA: understand frames, synchronous ticks and ego-camera placement.
  2. Reconstruct the plaza: read intrinsics, coordinate transforms, Gaussians and a held-out-view check.
  3. Generate a house: inspect meshes, materials and a camera path through real geometry.
  4. Generate a new WorldGen scene: compare text generation, observed image pixels and inferred geometry.
  5. Inspect a model's proposed driving trajectory, then evaluate a driving policy with AlpaSim.

Rendering a convincing scene does not supply a collision mesh or prove physical validity. Carry scale, coordinates and licensing metadata with the scene when connecting these tools. fVDB input/output theory · Media credits.