cs.AI

AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards

Xin Gao, Yifan Yang, Xin Li, Pan Wang, Yan Li, Muzhao Tian, Mingxi Cheng, Bei Liu, Kai Qiu, Qi Dai, Dong Chen, Yue Dong, Leandros Tassiulas, Rex Ying, Ji Li, Chong Luo

arXiv: 2610.18860Published: October 13, 202678 upvotes

View on arXiv Read PDF

AI summary

TL;DR

AesCode generates code for visual artifacts like slides and posters by using image generators for aesthetic guidance. It represents the target as a design graph and derives cross-modal rewards for optimization. AesCode improves spatial planning, aesthetic quality, and verifiable correctness.

Why It Matters

AesCode's approach enables the creation of high-quality, editable, and verifiable visual artifacts, which is crucial for various applications. This technology can be used in design, education, and presentation tools. The improvement in visual quality and correctness can enhance user experience and productivity.

Key Findings

AesCode improves visual quality by 31.1 and 22.4 percentage points
Reference images substantially improve visual quality
AesCode benefits from references as model input and visual rewards

Topics

reinforcement learning

multimodal

generative models

visual graph

cross-modal rewards

Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.