Overview
This project will study structured evaluation for 3D scene generation from multimodal and multi-view inputs.
Motivation
Visual fidelity alone does not capture geometric validity or spatial reasoning. A useful benchmark should make those dimensions independently inspectable.
Method
The planned workflow uses Blender-based scene inspection, multi-view evidence, and task-specific checks for geometry and spatial relationships.
Experiments and results
This page records a planned research direction. No experimental results or benchmark claims are presented yet.