AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards
AesCode uses generated reference images as aesthetic guidance while keeping code outputs checkable.
The paper frames visual code generation as a split problem: code is editable and verifiable, but weak at layout and color judgment, while image models are visually strong but unreliable for exact content. AesCode builds a design graph for the canvas, then scores generated code with separate rewards for verifiable properties and harder visual qualities judged by a VLM rubric. The authors say reinforcement learning over those decoupled rewards improves spatial planning, aesthetics, and correctness. Against same-scale Qwen3-VL baselines using the same references, AesCode-8B and AesCode-32B raise visual quality by 31.1 and 22.4 percentage points, respectively. ArXiv · AI/CL/LG's note
The paper frames visual code generation as a split problem: code is editable and verifiable, but weak at layout and color judgment, while image models are visually strong but unreliable for exact content. AesCode builds a design graph for the canvas, then scores generated code with separate rewards for verifiable properties and harder visual qualities judged by a VLM rubric. The authors say reinforcement learning over those decoupled rewards improves spatial planning, aesthetics, and correctness. Against same-scale Qwen3-VL baselines using the same references, AesCode-8B and AesCode-32B raise visual quality by 31.1 and 22.4 percentage points, respectively. ArXiv · AI/CL/LG's note
score 4