Creating visually appealing and well-structured slides, posters, or dashboards through AI-generated code is a promising way to produce editable and reliable designs. However, traditional code-generation models often struggle to capture how layout and color choices look together, limiting the overall aesthetic quality. Meanwhile, AI image generators can craft beautiful visuals but may not accurately represent the requested content, making it hard to translate their outputs into precise, editable code. A newly published research paper introduces AesCode, a novel approach that combines the strengths of both worlds by using AI-generated images as a guide to produce high-quality, verifiable design code. This innovation could pave the way for more efficient and visually pleasing automated design tools.
Key Takeaways
- AesCode generates design code guided by images created from the same content prompt, improving visual quality and layout planning.
- The system represents designs as a “design graph,” capturing all elements and their properties to enable detailed evaluation.
- It uses a combination of automated verifiers and a visual language model to assess both code correctness and aesthetic quality through separate reward signals.
- Compared to existing models, AesCode significantly boosts visual quality scores—up to 31 percentage points higher—matching or surpassing some leading closed-source AI systems.
The core challenge AesCode addresses is the disconnect between code generation and visual design quality. Code models typically generate structured outputs but lack “vision” to understand how these pieces fit visually. On the other hand, image generators excel at creating attractive visuals but can’t guarantee that the design can be reliably reproduced in code. To bridge this gap, the researchers designed AesCode to produce code that is both visually appealing and verifiably correct.
At the heart of AesCode is the concept of a “design graph,” a detailed representation that maps out every element on a canvas—such as shapes, colors, and spatial relationships—and their attributes. This graph allows the system to break down the design into components that can be checked for accuracy and aesthetic harmony. To evaluate the outputs, AesCode employs two types of “rewards.” First, deterministic verifiers analyze the code and its rendered image to score properties that can be directly parsed, ensuring correctness. Second, a Visual Language Model (VLM) acts as a judge for more subjective or complex visual qualities that are harder to quantify, using a custom Visual Graph Rubric tailored to the design elements and their relationships.
These rewards are “decoupled,” meaning they assess different aspects separately but are optimized together through reinforcement learning—a technique where the AI learns by receiving feedback from its performance. This approach helps the model improve both the spatial arrangement of design elements and the overall aesthetic appeal without sacrificing accuracy. The researchers also introduced a new benchmark to measure these qualities independently, finding that the visual scores align closely with human judgments.
One key insight from the study is the importance of reference images generated from the same prompt. Including these images as input and as a basis for visual rewards leads to substantial improvements in the generated code’s design quality. When compared with existing models of similar size, AesCode’s versions with 8 billion and 32 billion parameters achieved visual quality gains of 31.1 and 22.4 percentage points respectively. Remarkably, these results are on par with or even exceed those of some top-tier proprietary AI systems like GPT-5.5 and Claude Opus 4.8.
Looking ahead, AesCode’s approach could enhance tools that automatically generate editable, visually rich designs, benefiting industries like marketing, education, and data visualization. By ensuring that AI-generated visuals are both beautiful and reliably reproducible as code, this research opens the door to more creative and efficient workflows. Future work may explore extending these methods to more complex design types or integrating user preferences more deeply, further bridging the gap between AI’s creative and technical capabilities.
Based on research published on arXiv by Xin Gao, Yifan Yang, Xin Li et al..