Generative Blocks World Enables 3D-Primitive-Based Interactive Editing of Generated Images

Researchers from the University of Illinois have introduced Generative Blocks World, a system that lets users interactively edit generated images by representing scenes as assemblies of convex 3D primitives. Editors can move whole structures or fine details by manipulating these geometric abstractions, after which a flow-based image generation method re-renders the scene conditioned on depth maps and texture hints derived from the modified primitives.

The texture-hint mechanism is central to the approach: by accounting for the altered 3D geometry, it preserves object identity and compositional consistency across both object and camera moves — outperforming prior methods on visual fidelity and editability, according to the authors. The work was accepted at ICLR 2026.