Structured 3D Latents Are Surprisingly Powerful: Unleashing Generalizable Style with 2D Diffusion

Yiran Qiao, Yiren Lu, Yunlai Zhou, Disheng Liu, Linlin Hou, Rui Yang, Yu Yin, Jing Ma · 2026

Read Structured 3D Latents Are Surprisingly Powerful: Unleashing Generalizable Style with 2D Diffusion by Yiran Qiao, Yiren Lu, Yunlai Zhou, Disheng Liu, Linlin Hou, Rui Yang, Yu Yin, Jing Ma (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-04412.txt.

Description

Yiran Qiao and seven co-authors present a method for generating three-dimensional objects in a chosen visual style, for uses such as games and virtual reality. Existing methods work when the target style resembles their training data, but degrade badly on unfamiliar styles.

Their approach uses a pretrained two-dimensional diffusion model as a teacher, since such models carry broad and transferable style knowledge. Rendered views of the object are aligned with the target style under that guidance, and the structured three-dimensional representation is adjusted to match. The authors argue the bottleneck was never model capacity, but how little past methods used those structured representations.

Catalog record

CREATORYiran Qiao, Yiren Lu, Yunlai Zhou, Disheng Liu, Linlin Hou, Rui Yang, Yu Yin, Jing Ma
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.04412

Browse the full library · About the Tower of Babel