A research pipeline that walks users through a conversation before generating any geometry outperformed open-ended prompt tools in a controlled VR study.
CoGen3D is an agentic pipeline built to lower the barrier to 3D asset creation for virtual reality. Instead of dropping users in front of a blank prompt box, it guides them through intent elicitation — a structured back-and-forth — then shows a concept image for approval before committing to full 3D generation and scene deployment. Researchers tested the system with 120 participants across six emotionally varied immersive scenes: 60 co-created assets, 60 evaluated the finished scenes.
The engagement numbers favored the co-design approach, and participants' emotional responses shifted in ways the researchers tied to the collaborative process. The wrinkle: users consistently preferred the concept images over the final 3D models, and they did not go easier on their own creations — the study found no evidence of the "IKEA effect" inflating self-assessments of generated work. That preference gap is a quiet indictment of current image-to-3D conversion quality, which remains the pipeline's weakest link.
The broader pattern here is familiar — conversational scaffolding consistently outperforms cold prompting when non-expert users need to externalize fuzzy ideas — but applying it end-to-end in a VR authoring context is a genuine step. The real test will be whether image-to-3D fidelity catches up to the expectations the concept-image stage sets.