Roblox Corporation, a global platform for immersive user-generated content, has introduced Cube 3D, a foundational generative AI model for creating 3D digital content from text prompts. The company has made a version of the model open-source, now accessible via GitHub and HuggingFace. The beta version, which integrates into Roblox Studio and its Lua-based API, enables creators to produce 3D meshes directly within game environments using natural language input.
Unlike image-based reconstruction methods that rely on limited visual data, Cube 3D is trained on native 3D assets generated and used across Roblox’s ecosystem. This allows it to output structurally complete digital objects compatible with game engines—objects that can be interacted with in gameplay, rather than being flat facades. A typical use case involves entering a command like “generate a motorcycle” into the platform’s Assistant, resulting in a full 3D mesh suitable for immediate in-game deployment. These objects can later be enhanced with texture and color but are generated as functionally usable meshes from the outset.
Cube 3D applies a token-based system to understand and predict 3D shapes. Drawing from techniques used in large language models, the AI converts geometry into shape tokens and uses autoregressive transformers to forecast subsequent tokens in a sequence—effectively “building” a mesh piece by piece. This method supports both individual object completion and full scene layout generation. To align multimodal inputs, Roblox engineers developed a unified transformer architecture compatible with text, image, and future data types such as audio. The current release focuses on object generation from text, but future updates are expected to support scene-level outputs and hybrid input modalities.
Roblox has positioned the software as part of a broader shift toward real-time, user-augmented content creation. Players and developers alike will be able to generate props, environments, or interactive objects on demand. The longer-term aim is “4D creation,” where AI understands not only object form but also interaction logic and environmental relationships. This includes bounding box placement for layout, mesh fusion for multi-object environments, and context-aware alterations—such as swapping seasonal elements or adapting geometry based on narrative triggers within a game.
While Cube 3D does not currently support 3D printing file formats such as STL, the underlying methodology of tokenizing 3D shapes may influence emerging tools in virtual prototyping, AI-assisted design, and even CAD automation. Open-source release of this nature remains rare among proprietary game development platforms, particularly those dealing with native 3D asset pipelines. Roblox has also co-founded ROOST, a nonprofit focused on open-source AI safety, and recently released other models tied to responsible AI development.
Emerging AI-generated 3D modeling
Tencent, a Chinese multinational technology company with significant investments in gaming and cloud services, released Hunyuan3D 2.0 to streamline digital asset creation. The system features two specialized models—Hunyuan3D-DiT for geometry and Hunyuan3D-Paint for texture—designed to improve fidelity and responsiveness in 3D generation. Internal benchmarks using Condition-Model Matching Distance (CMMD) and Fréchet Inception Distance (FID) suggest higher alignment between user prompts and resulting outputs. A companion interface, Hunyuan3D-Studio, enables sketch-to-3D workflows and low-polygon mesh export. While Tencent has not formally targeted additive manufacturing, the platform’s ability to generate both high-resolution and simplified meshes may support adaptation in prototyping and multi-material 3D printing environments.
A year earlier, Nvidia, a leading developer of GPUs and parallel computing platforms widely used in AI and visualization, introduced Magic3D in 2023 as part of its generative AI research. The tool produces textured 3D meshes from natural language using a two-stage pipeline that refines coarse models into high-resolution geometry. Demonstrations—such as rendering a blue poison dart frog from a single prompt—illustrated the system’s capacity for mesh synthesis, text-driven editing, and stylistic transformation. Although primarily geared toward video games and computer-generated imagery, Nvidia researchers identified potential crossover applications in VR development and special effects, noting that streamlined model generation could lower production barriers across digital design domains.
Ready to discover who won the 2024 3D Printing Industry Awards?
Subscribe to the 3D Printing Industry newsletter to stay updated with the latest news and insights.
Featured image showcase Prompt: A red buggy with knobby tires. Image via Roblox.
