Prompted Props, Human Pipelines: Evaluating AI-Generated 3D Assets for Game-Ready Environments

Résumé fourni par la source

Abstract: Generative AI systems increasingly promise rapid 3D asset production for game development, yet their practical viability depends on whether generated models can move beyond visual preview into editable, optimized, engine-ready workflows. This article presents a practice-led comparative case study of a stylized fantasy tavern environment produced through two workflows: a human-authored Blender pipeline and an AI-assisted pipeline using Meshy 6 and Hunyuan 3D. Using a fixed asset list, shared visual theme, documented prompts, production-time tracking, visual comparison, topology inspection, UV-map analysis, and post-generation labor accounting, the study evaluates whether text-to-3D tools function as production substitutes, ideation accelerators, or conditional asset sources. Results indicate that AI-assisted generation substantially reduced first-pass production time: the Hunyuan-assisted reconstruction required 238 minutes compared with 716.06 minutes for the human-authored scene. However, the apparent time savings were accompanied by substantial technical debt, including dense triangulated geometry, fragmented UV maps, inconsistent prompt adherence, material-editing constraints, clipping during placement, and loss of texture integrity during attempted decimation. The human-authored workflow required more labor at the modeling stage but produced assets with greater intentionality, editability, scale control, and technical legibility. The findings support a human-in-the-loop model in which generative AI contributes most effectively to ideation, rough prop exploration, and early prototyping, while artists and technical artists remain necessary for optimization, art direction, retopology, UV reconstruction, material refinement, and engine validation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.