Square Enix Details ATOGiS 3D Generation AI at CEDEC2026
Square Enix built a 3D generation AI trained only on its own assets, pairing a TRELLIS.2-based geometry model with a proprietary staged retopology architecture, to put its rarely reused asset library to work.
Reporting from 1 source: GameBusiness.jp.
Square Enix presented ATOGiS, its proprietary 3D generation AI, at CEDEC2026. The system combines an image-to-3D geometry AI with a retopology AI, both trained on roughly 60,000 in-house weapon and background assets to avoid licensing problems. AI & Engine Development Division member Taketo Araki led the session, covering dataset selection, a staged retopology learning method, and three pre-tests run before large-scale training on eight H100 GPUs.
Square Enix's 3D assets are designed to fit each work's world view, so they rarely get reused. To make use of them, the company built its own generation AI. The ATOGiS dataset came to about 60,000 in-house weapon and background assets, with complex structures, textures, and externally licensed items excluded. Polygons over 10,000 were cut to limit GPU load.
The retopology AI learns in stages: vertices first, then edges, then faces. Vertices are converted to voxel grids, edges are guided by a matrix, and faces are filled by a standard algorithm. The geometry AI customizes the TRELLIS.2 architecture, converting input images to normal maps and adding a geometry normal loss to speed learning. Before large-scale training, the team ran three tests: a learning expression test, a single overfitting test, and a category learning test.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.
Sources
- GameBusiness.jp 自社3Dアセットで学習した3D生成AIを研究開発する…スクウェア・エニックスの取り組みとは？【CEDEC2026】