AI 3D scene creation tools
Updated 2026-10-12
3D World & Scene Generation
A multimodal world model that generates explorable, persistent 3D worlds from text, images, video, or coarse 3D layouts, with tools to edit, expand, and combine generated worlds.
- Marble became generally available on November 12, 2025, accepting text, single or multiple images, video, and coarse 3D layouts (via the Chisel tool) as input. [worldlabs.ai]
- Generated worlds can be exported as Gaussian splats, triangle meshes (including collider meshes), or videos with camera control.
- World Labs announced the World API on January 21, 2026, exposing Marble's world generation programmatically via platform.worldlabs.ai, with early integrations including NVIDIA Isaac Sim and MuJoCo. [worldlabs.ai]
An open-source 3D world generation framework that turns text or images into immersive, explorable panoramic worlds with exportable, semantically layered 3D meshes.
- HunyuanWorld-1.0, released July 26, 2025, uses a semantically layered 3D mesh representation with panoramic images as 360° world proxies, and outputs meshes compatible with standard graphics pipelines.
- Follow-up releases include HunyuanWorld-1.0-lite for consumer GPUs (August 2025), WorldMirror for video/multi-view input (October 2025), and HunyuanWorld-1.5 'WorldPlay' for real-time world creation (December 18, 2025).
- HY-World 2.0 was announced on April 16, 2026, with model weights distributed via Hugging Face.
Interactive World Models — Video Simulation
An AI lab building general world models that stream long-form interactive video simulations from text or image prompts, explorable in real time from the browser.
- Odyssey-2, the company's flagship general-purpose world model, generates multi-minute interactive simulations with streaming that initiates in about 50ms, accessible at experience.odyssey.ml.
- Agora-1 is a multi-agent world model enabling simultaneous real-time interaction between human and AI participants in shared simulations.
- Odyssey raised a $310M Series B; TechCrunch reported the round in June 2026 at a $1.45B valuation with Amazon among the backers. [techcrunch.com]
A real-time AI lab offering Oasis, an API-accessible interactive world model aimed at physical AI (robotics, autonomous vehicles), and Lucy, a real-time video/world editing model.
- Decart describes Oasis 3 as an API-accessible world model for physical AI that generates interactive, physically accurate environments in real time, targeting robotics, autonomous vehicles, manufacturing, and drones.
- Lucy 2.0 performs real-time, text-driven video transformation at production scale, and MirageLSD is presented as achieving infinite real-time video generation.
- Developer access, documentation, and pricing are provided through Decart's API platform at platform.decart.ai.
An interactive world model that generates navigable environments as video in real time. Google links to Project Genie, an experimental research prototype, from its model page.
- Google describes real-time interaction at 20–24 frames per second and 720p output.
- The environments are generated frame by frame from a world description and user actions; this differs from an exported mesh or Gaussian-splat scene.
- The official model page links to Project Genie and describes it as an experimental research prototype. Check that service for current access requirements.
3D Asset Generation
A web-based AI 3D generator that produces textured 3D models from text prompts or images, with AI texturing, auto-rigging, and animation for game and 3D-printing workflows.
- Text-to-3D generates a textured model from a written prompt in the browser. [meshy.ai]
- Documented exports include FBX, OBJ, GLB, USDZ, STL, BLEND and 3MF; check the export interface for the formats available to your generation.
- Post-generation tools include texture generation, remeshing, a model viewer and printability checks.
A text-to-3D and image-to-3D platform with APIs for texturing, mesh editing and animation. Its official API documentation also links to the Tripo Studio web application.
- Tripo’s H3 API supports v3.0 and v3.1 models, with Standard geometry up to 1.5 million triangles and Ultra geometry up to 2 million. [docs.tripo3d.ai]
- The texture API supports PBR material options and texture regeneration; texture-quality settings and compatible input versions affect the supported operations. [docs.tripo3d.ai]
- Official documentation provides a REST API and Python SDK for submitting generation tasks and retrieving completed results. [docs.tripo3d.ai]
- API usage is prepaid and credit-based at $1 per 100 credits; generation and optional post-processing tasks have separate credit costs. These API prices are distinct from Studio subscriptions.
An AI 3D model generator that creates 3D meshes from text prompts or images, with mask-based 3D editing (inpainting) of existing models.
- Rodin Gen 2 generates 3D models from text or images and supports 3D inpainting that uses a masked diffusion process to modify selected regions of a model. [hyper3d.ai]
- Outputs high-poly quad meshes positioned as production-ready, with export to STL, FBX, OBJ, GLB, and USDZ.
- Rodin image-to-3D generation is also offered as a hosted API through fal.ai. [fal.ai]
A game-asset generator combining parametric templates with AI, producing customizable, 3D models from text prompts, images, or slider-based template editing.
- Generates 3D models from text prompts, photos, sketches, and parametric templates with adjustable controls, under a subscription model with unlimited generation rather than per-credit consumption.
- Offers topology control with adjustable polygon counts, one-click auto-rigging, AI retexturing, and manifold geometry suitable for 3D printing.
- Available as a web app with plugins for Unity, Unreal, and Blender, and includes a Roblox avatar creator.
Roblox's open-source generative AI system for 3D creation, generating meshes from text prompts inside Roblox Studio and via a Lua API, with a roadmap toward scene generation.
- Cube launched March 17, 2025 with text-to-mesh generation in Roblox Studio and an in-experience Lua API (e.g. '/generate a motorcycle').
- The model is open-sourced on GitHub and Hugging Face, and developers can fine-tune it or train it on their own data.
- Cube tokenizes 3D shapes the way language models tokenize text, predicting shape tokens autoregressively; Roblox's stated roadmap extends to scene generation and '4D creation'.
An open-source large 3D asset generation model using a unified Structured LATent (SLAT) representation, decoding to radiance fields, 3D Gaussians, or meshes from text or image prompts.
- TRELLIS outputs Radiance Fields, 3D Gaussians, and meshes from image or text prompts, with local 3D editing and variant generation.
- Model lineup ranges from 342M to 2.0B parameters (image-large: 1.2B; text-xlarge: 2.0B), trained on the 500K-object TRELLIS-500K dataset.
- The models and main codebase are MIT-licensed; the image model shipped December 2024 and training code plus text models followed on March 25, 2025.
Skyboxes & Panoramic Backdrops
A text-to-360° environment generator that produces seamless equirectangular skybox panoramas for games, VR, and virtual production, with API/SDK and engine export paths.
- Generates equirectangular 360° panoramas at up to 16K resolution, with 32-bit HDRI export for scene lighting.
- Offers an API and SDK, a native Unity plugin, direct exports to Blender and Unreal, and MCP (Model Context Protocol) server support.
Capture & 3D Reconstruction
A capture platform that turns phone photos and video of real objects and places into photorealistic, embeddable interactive 3D scenes viewable on the web.
- Available on web, iOS, and Android, with scenes that stream immediately and run at 30 FPS in web browsers.
- Captures are usable commercially without additional licensing and remain private unless made public; a Luma API supports bulk generation.
- Luma publishes an open web library and three.js examples for embedding its captures in WebGL projects. [github.com]
A reality-capture platform for existing objects and spaces, with spatial, object, floor-plan and aerial capture workflows.
- Polycam documents AI-assisted photogrammetry and LiDAR capture for real objects and places. LiDAR requires compatible hardware.
- Its site lists spatial capture, object capture, floor plans, drone/aerial workflows and Gaussian-splat tools.
Epic Games' photogrammetry software (formerly RealityCapture) that reconstructs 3D models from photos and laser scans, with desktop and mobile versions.
- Free for individuals, educators, and companies under $1 million USD annual gross revenue; paid seats apply above that threshold.
- Supports photogrammetry plus laser-scan input, SLAM data import, and classified point clouds; version 2.2 added full AMD GPU support (Radeon, RDNA 4, Ryzen AI Max).
- Ships as a Windows desktop app via the Epic Games Launcher plus a RealityScan Mobile app for iOS/Android.
A freemium 3D scanning app for phone and web that reconstructs objects and spaces via photogrammetry, LiDAR, NeRF-based featureless-object scanning, and 3D Gaussian splatting.
- Offers Photo Scan (photogrammetry), LiDAR Scan (on supported iOS devices), featureless-object scanning using NeRF, and 3D Gaussian splatting with mesh export.
- Available on iOS, Android, and web browsers under a freemium model.
- Supports quad mesh output with PBR materials, mesh decimation, and auto-rigging.
A Windows desktop application for training radiance-field/Gaussian-splatting scenes locally from photos or video, with live training preview and editing tools.
- Builds 3D scenes from regular RGB images using radiance field technology, with all processing done locally on the user's machine.
- Requires Windows 10+ and an NVIDIA GPU with Compute Capability 7.5 or higher (e.g. GeForce RTX 2060 or better).
- Features include live training preview, scene merging, image masking, AprilTag alignment, GUI and command-line interfaces, and Unreal Engine and After Effects integrations.
Engine & Scene-Assembly Tools
An AI assistant for environment artists that assembles scenes from natural-language requests using a team's own asset libraries, working inside standard 3D tools and engines.
- Provides AI world building from natural-language requests and reasons over a team's existing creative assets (images, video, 3D models, animations) via a local metadata library.
- Ships open-source plugins for Unreal Engine (the primary integration), Unity, 3ds Max, Maya, and Blender.
- Asset files stay on the user's systems; the company states it does not upload, store, or train on customer assets.
An AI toolkit plugin for Unreal Engine that generates C++ code, Blueprints, and 3D models, and assists with scene creation and engine-specific questions.
- Generates C++ code, 3D models, and functional Blueprints inside Unreal Engine, with documentation at docs.ludusengine.com.
- Supports Unreal Engine versions 5.4 to 5.7 and offers a free trial.
Unity's in-editor AI suite for Unity 6, combining an agentic Assistant, asset Generators, an AI Gateway, and an MCP server for editor automation and asset creation.
- Unity AI's open beta was announced May 1, 2026, including an agentic in-project Assistant, Generators, an AI Gateway, and a Model Context Protocol server. [discussions.unity.com]
- Generators create placeholder materials, sounds, cubemaps, and 2D/3D assets in-editor.
- Personal Edition users get a 14-day trial with 1,000 one-time credits, then $10/month for 1,000 monthly credits; Pro/Enterprise/Industry seats include access. Unity states it does not train on user code or data by default.
- The Generators package (com.unity.ai.generators) is distributed for Unity 6000.2+ and was in pre-release (1.0.0-pre.20) as of this documentation version. [docs.unity3d.com]
NVIDIA's platform of libraries and services for building OpenUSD-based 3D pipelines, industrial digital twins, and physical-AI simulation, with generative AI microservices for scene search and assembly.
- USD Code and USD Search NIM microservices are generally available, letting developers use text prompts to generate or search for OpenUSD assets (announced at CES, January 6, 2025). [nvidianews.nvidia.com]
- NVIDIA Edify SimReady automatically labels existing 3D assets with attributes such as physics and materials, processing 1,000 objects in minutes versus roughly 40 hours manually.
- Omniverse blueprints announced alongside include Mega (robot-fleet testing in industrial digital twins), AV simulation, spatial streaming to Apple Vision Pro, and real-time digital twins for CAE.
Research Papers — Methods & Prototypes
Research paper: a research system that converts a single indoor panorama into a persistent, explorable 3D Gaussian scene for free-viewpoint navigation.
- Pano2World directly outputs a persistent, explorable 3D Gaussian scene from a single panorama, using coarse 3D Gaussian reconstruction and a panoramic diffusion model with View-Aware Attention Routing.
- The paper was submitted to arXiv on July 1, 2026 and targets indoor panorama-to-3D generation for free-viewpoint navigation.
Research paper: a text-driven 3D indoor scene synthesis framework for non-Manhattan environments, focused on physically plausible object layouts in rooms with non-orthogonal geometry.
- SPG-Layout uses statistical priors of object distributions and a hierarchical layout strategy to generate physically plausible indoor scenes in complex non-Manhattan environments.
- The paper introduces a benchmark of 500 diverse non-Manhattan environments for evaluating text-driven indoor scene synthesis.
Research paper: an open-source framework for long-horizon, playable video world generation, covering data preparation, model architecture, training, inference acceleration and deployment for interactive generative worlds.
- AlayaWorld is presented as a full-stack open-source framework for interactive generative worlds, enabling real-time navigation and actions such as combat, spell casting and monster summoning.
- The paper frames playable world generation as autoregressive synthesis of future observations conditioned on current world state and user interactions, trained on gameplay recordings and real-world videos.
Research paper: a multi-agent system for creating dynamic, editable 4D scenes from text in Blender, coordinating layout, physics solvers, temporal sequencing, camera and lighting.
- SimWorlds uses a planner-coder-reviewer multi-agent workflow plus a deterministic verifier to produce dynamic 4D scenes with liquids, particles, rigid bodies and articulated mechanisms.
- The paper introduces 4DBuildBench, a benchmark for assessing dynamic text-to-scene creation beyond static 3D scene output.
Research paper: a research framework for scene-scale 3D diffusion that turns prompt-driven dimetric scene images into globally coherent, layout-controllable 3D scenes of arbitrary size.
- SynCity 3000 extends image-to-3D generation to entire scenes by adapting a generator into a convolutional operator and fine-tuning it on scene-like data from a synthetic data engine.
- The authors report that the method produces large, coherent and detailed 3D scenes from diverse prompts and layouts, addressing prior limitations in scene-scale 3D generation.
Research paper: a unified pixel-space diffusion model for 3D scene generation and reconstruction, optimizing directly on rendered images instead of latent 3D features.
- PixWorld reformulates 3D reconstruction and 3D generation under one pixel-space diffusion paradigm, supervising diffusion on rendered images to better align optimization with scene fidelity.
- The paper was submitted to arXiv on July 6, 2026 and targets a single model that handles both 3D scene reconstruction and generation.
Research paper: a unified 3D foundation-model architecture that uses elastic semantic anchoring to connect language cues with geometry at multiple abstraction scales.
- ELSA3D introduces Anchor Tokens that select semantic cues, route them to relevant 3D scales, retrieve scale-specific geometric evidence, and write the fused signal back into a unified 3D representation.
- The paper was submitted to arXiv on July 7, 2026 and focuses on unified 3D understanding and generation rather than only asset synthesis.
Research paper: a research framework for real-world panoramic world generation that addresses long-range memory in omnidirectional world models and introduces the World360 panoramic video dataset.
- PanoWorld simplifies camera trajectories into translations with fixed headings and uses Dense Panoramic Ray-Conditioning plus Geometry-aware Memory Augmentation for long-range panoramic world modeling.
- The paper introduces World360, a large-scale dataset of real-world panoramic UAV video clips and synthetic panoramic video clips for evaluating physical consistency under large spatial and illumination changes.
Research paper: a generative, simulation-ready 3D world engine for embodied AI that assembles policy-ready environments from assets, affordances, tasks and multi-room scenes.
- EmbodiedGen V2 presents a unified sim-ready representation connecting cross-simulator assets, interaction affordances, task-driven worlds, large-scale multi-room scenes and editable stateful environments.
- The paper states that generated environments support manipulation, navigation, mobile manipulation and cross-simulator deployment, aiming to reduce manual assembly in closed-loop embodied-AI learning pipelines.
Historical References
Historical reference: Meta documented CSM’s use of image/video inputs and SAM 2 segmentation to produce 3D assets. The former CSM website and service domains were unreachable during the October 11, 2026 check; current hosted-service availability is unconfirmed.
- Meta’s case study described CSM converting images and video into segmented 3D assets prepared for rigging and animation, for use in game engines, VR and VFX.
- The case study described CSM using Meta’s Segment Anything Model 2 (SAM 2) to segment elements in images and video before producing 3D components.