InfiniDream: Open-Ended Photorealistic 3D Worlds Through Procedural Generation

InfiniDream. Open-Ended Photorealistic 3D Worlds Through Procedural Generation. By Yi-Chuan Huang and five others, from Alaya Lab, National Yang Ming Chiao Tung University, and one other institution. Published on the nineteenth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Procedural generation plays a key role in synthetic data creation and in game development. Compared with static mesh representations, it offers greater controllability, editability and diversity. The difficulty is getting good procedural content in the first place. VLM-based code generation often produces relatively simple 3D assets. Sophisticated photorealistic procedural systems do exist, but they require extensive manual development. And existing VLM-guided procedural methods remain largely constrained by predefined factories and APIs, which limits their ability to create novel asset types. Each option gives up either quality, effort or range.

InfiniDream is a VLM-driven framework for open-ended procedural generation of 3D worlds. It makes two contributions, one at the level of individual assets and one at the level of whole scenes, and together they run from a text prompt to a populated, animated world.

For assets, Symbolically Constrained Asset Generation leverages existing procedural algorithms to create novel, photorealistic assets, while preserving editability and seed-based diversity. For worlds, an end-to-end text-to-world pipeline combines semantic scene planning with Deterministic Frustum-Guided Grid Placement. That combination constructs large-scale 3D worlds with precise object placement, procedural population, and animation. Planning works at the level of meaning, deciding what a scene should hold. Placement works at the level of geometry, and it is deterministic, so the same plan places its objects the same way each time.

Extensive evaluations on one hundred object generation tasks and twenty scene generation tasks show improvements across four perceptual metrics. Diversity analysis, ablation studies, and human preference evaluations further support the effectiveness of the framework, checking the variety of its output, the contribution of its parts, and how people judge the results.

Rather than writing assets from scratch as code, or choosing among predefined factories, InfiniDream builds on procedural algorithms that already exist, and that is how it aims for assets that are both novel and photorealistic. Seeds and editability are kept throughout, so the output stays procedural content rather than settling into a static mesh.