generated-3d-worlds technique sourced
Prompt-to-persistent 3D world generation
Models such as World Labs Marble and HunyuanWorld turn text, images, video, or coarse 3D blockouts into persistent explorable scenes exported as Gaussian splats and meshes with collision proxies.
mechanism
An image or video diffusion model generates views of a scene (a 360-degree panorama in HunyuanWorld, multi-view images in Marble) from the prompt; the views are lifted to an explicit 3D representation (3D Gaussian splats, or a semantically layered mesh) that can be walked through and re-rendered without regeneration. Marble also emits triangle meshes and coarse collider meshes for physics, can expand or merge worlds, and in Chisel mode takes a blockout of boxes and planes as structure and a text prompt as style. A parallel line renders frames directly instead of building geometry: RTFM keeps generated frames with camera poses as spatial memory and retrieves nearby frames as context. Atlas (2026) is a multimodal model that outputs video, point clouds, and splats.
limitations
- Scenes are static: no dynamic objects, agents, or rules are generated, and RTFM lists dynamics and interaction as future work.
- Single-image or text prompts force the model to invent unseen regions; Marble notes artifacts and less crisp far regions that expansion may fix.
- Splat output bakes lighting into colors, so day-night cycles and relighting need conversion to meshes or other methods.
- Collider meshes are coarse and need manual work for gameplay-critical collision.
- Corporate change (World Labs joining AMD in 2026) leaves the hosted product's future terms uncertain.
appears in
the paper, 3 sections · 1 chapter · 1 episoderelations
5 stored edges
Procedural build assist
Chisel turns a coarse blockout into a detailed styled scene, the same rough-input to detailed-output contract as Townscaper-style build assist but with a learned generator; today it is an editor mode, not a runtime player tool.
Procedural terrain
Maps come from a prompt, photo, or video rather than a seed and noise pipeline, and unlike frame-based world models the result is baked geometry that persists, can be saved, and can be expanded region by region.
Rigid-body physics
Exported collider meshes give generated scenes collision geometry automatically; Marble describes them as coarse, so precise physics still needs cleanup.
Scenario objectives
Bespoke scenario maps can be produced from a reference image or description and exported with collider meshes, reducing hand modeling per scenario; objectives and rules still need authoring.
claims
4 with sourcesMarble became generally available on 2025-11-12, taking text, images, video, or coarse 3D layouts and exporting Gaussian splats, triangle meshes including collider meshes for coarse physics, and rendered video; Chisel separates blocked-out structure from prompted style.
HunyuanWorld 1.0 (Tencent, July 2025) generates explorable worlds from text or images via panoramic proxies converted to a semantically layered 3D mesh exportable to standard graphics pipelines, listing game development as an application.
RTFM (October 2025) renders generated or real scenes from a single image at interactive frame rates on one H100, using posed frames as spatial memory; it models static scenes only.
World Labs launched a public World API in January 2026, released Atlas to early-access partners in September 2026, and announced it is joining AMD on 2026-09-28.